minimize fill_

Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>

.buildkite/check-wheel-size.py

@@ -1,12 +1,14 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import os
 import sys
 import zipfile
 
-# Read the VLLM_MAX_SIZE_MB environment variable, defaulting to 300 MiB
+# Read the VLLM_MAX_SIZE_MB environment variable, defaulting to 400 MiB
 # Note that we have 400 MiB quota, please use it wisely.
 # See https://github.com/pypi/support/issues/3792 .
 # Please also sync the value with the one in Dockerfile.
-VLLM_MAX_SIZE_MB = int(os.environ.get('VLLM_MAX_SIZE_MB', 300))
+VLLM_MAX_SIZE_MB = int(os.environ.get('VLLM_MAX_SIZE_MB', 400))
 
 
 def print_top_10_largest_files(zip_file):

.buildkite/generate_index.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import os
 

.buildkite/lm-eval-harness/test_lm_eval_correctness.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 LM eval harness on model to compare vs HF baseline computed offline.
 Configs are found in configs/$MODEL.yaml

.buildkite/nightly-benchmarks/scripts/convert-results-json-to-markdown.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import json
 import os
 from pathlib import Path

.buildkite/nightly-benchmarks/scripts/download-tokenizer.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 
 from transformers import AutoTokenizer

.buildkite/nightly-benchmarks/scripts/generate-nightly-markdown.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import json
 from pathlib import Path

.buildkite/nightly-benchmarks/scripts/get-lmdeploy-modelname.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from lmdeploy.serve.openai.api_client import APIClient
 
 api_client = APIClient("http://localhost:8000")

.buildkite/nightly-benchmarks/scripts/summary-nightly-results.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import datetime
 import json
 import os

.buildkite/run-gh200-test.sh

@@ -23,6 +23,6 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Run the image and test offline inference
-docker run --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
-    python3 examples/offline_inference/basic.py
+docker run -e HF_TOKEN -v /root/.cache/huggingface:/root/.cache/huggingface --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
+    python3 examples/offline_inference/cli.py --model meta-llama/Llama-3.2-1B
 '

.buildkite/test-pipeline.yaml

@@ -50,9 +50,9 @@ steps:
   - tests/multimodal
   - tests/test_utils
   - tests/worker
-  - tests/standalone_tests/lazy_torch_compile.py
+  - tests/standalone_tests/lazy_imports.py
   commands:
-  - python3 standalone_tests/lazy_torch_compile.py
+  - python3 standalone_tests/lazy_imports.py
   - pytest -v -s mq_llm_engine # MQLLMEngine
   - pytest -v -s async_engine # AsyncLLMEngine
   - NUM_SCHEDULER_STEPS=4 pytest -v -s async_engine/test_async_llm_engine.py
@@ -349,6 +349,7 @@ steps:
   - vllm/
   - tests/models
   commands:
+    - pytest -v -s models/test_transformers.py
     - pytest -v -s models/test_registry.py
     - pytest -v -s models/test_initialization.py
 
@@ -485,6 +486,7 @@ steps:
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
   - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
   # Avoid importing model tests that cause CUDA reinitialization error
+  - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
   - pytest models/encoder_decoder/language/test_bart.py -v -s -m 'distributed(num_gpus=2)'
   - pytest models/encoder_decoder/vision_language/test_broadcast.py -v -s -m 'distributed(num_gpus=2)'
   - pytest models/decoder_only/vision_language/test_models.py -v -s -m 'distributed(num_gpus=2)'

.pre-commit-config.yaml

@@ -97,10 +97,14 @@ repos:
     language: system
     verbose: true
     stages: [commit-msg]
+  - id: check-spdx-header
+    name: Check SPDX headers
+    entry: python tools/check_spdx_header.py
+    language: python
+    types: [python]
   - id: suggestion
     name: Suggestion
     entry: bash -c 'echo "To bypass pre-commit hooks, add --no-verify to git commit."'
     language: system
     verbose: true
     pass_filenames: false
-

CODE_OF_CONDUCT.md

@@ -61,7 +61,7 @@ representative at an online or offline/IRL event.
 
 Instances of abusive, harassing, or otherwise unacceptable behavior may be
 reported to the community leaders responsible for enforcement in the #code-of-conduct
-channel in the [vLLM Discord](https://discord.com/invite/jz7wjKhh6g).
+channel in the [vLLM Slack](https://slack.vllm.ai).
 All complaints will be reviewed and investigated promptly and fairly.
 
 All community leaders are obligated to respect the privacy and security of the

Dockerfile

@@ -127,7 +127,7 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
 # Check the size of the wheel if RUN_WHEEL_CHECK is true
 COPY .buildkite/check-wheel-size.py check-wheel-size.py
 # sync the default value with .buildkite/check-wheel-size.py
-ARG VLLM_MAX_SIZE_MB=300
+ARG VLLM_MAX_SIZE_MB=400
 ENV VLLM_MAX_SIZE_MB=$VLLM_MAX_SIZE_MB
 ARG RUN_WHEEL_CHECK=true
 RUN if [ "$RUN_WHEEL_CHECK" = "true" ]; then \

README.md

@@ -10,7 +10,7 @@ Easy, fast, and cheap LLM serving for everyone
 </h3>
 
 <p align="center">
-| <a href="https://docs.vllm.ai"><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> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
+| <a href="https://docs.vllm.ai"><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://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>
 
 ---
@@ -36,7 +36,7 @@ Easy, fast, and cheap LLM serving for everyone
 ## About
 vLLM is a fast and easy-to-use library for LLM inference and serving.
 
-Originally developed in the [Sky Computing Lab](https://sky.cs.berkeley.edu) at UC Berkeley, vLLM has evloved into a community-driven project with contributions from both academia and industry.
+Originally developed in the [Sky Computing Lab](https://sky.cs.berkeley.edu) at UC Berkeley, vLLM has evolved into a community-driven project with contributions from both academia and industry.
 
 vLLM is fast with:
 
@@ -139,8 +139,7 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
 ## Contact Us
 
 * For technical questions and feature requests, please use Github issues or discussions.
-* For discussing with fellow users, please use Discord.
-* For coordinating contributions and development, please use Slack.
+* For discussing with fellow users and coordinating contributions and development, please use Slack.
 * For security disclosures, please use Github's security advisory feature.
 * For collaborations and partnerships, please contact us at vllm-questions AT lists.berkeley.edu.
 

benchmarks/backend_request_func.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import json
 import os
 import sys

benchmarks/benchmark_guided.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """Benchmark guided decoding throughput."""
 import argparse
 import dataclasses

benchmarks/benchmark_latency.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """Benchmark the latency of processing a single batch of requests."""
 import argparse
 import dataclasses

benchmarks/benchmark_long_document_qa_throughput.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 Offline benchmark to test the long document QA throughput.
 

benchmarks/benchmark_prefix_caching.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 Benchmark the efficiency of prefix caching.
 

benchmarks/benchmark_prioritization.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """Benchmark offline prioritization."""
 import argparse
 import dataclasses

benchmarks/benchmark_serving.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 r"""Benchmark online serving throughput.
 
 On the server side, run one of the following commands:

benchmarks/benchmark_serving_guided.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 r"""Benchmark online serving throughput with guided decoding.
 
 On the server side, run one of the following commands:

benchmarks/benchmark_throughput.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """Benchmark offline inference throughput."""
 import argparse
 import dataclasses

benchmarks/cutlass_benchmarks/sparse_benchmarks.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import copy
 import itertools

benchmarks/cutlass_benchmarks/utils.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 # Cutlass bench utils
 from typing import Iterable, Tuple
 

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import copy
 import itertools

benchmarks/cutlass_benchmarks/weight_shapes.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 # Weight Shapes are in the format
 # ([K, N], TP_SPLIT_DIM)
 # Example:

benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import os
 
 import aiohttp

benchmarks/disagg_benchmarks/round_robin_proxy.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import asyncio
 import itertools
 

benchmarks/disagg_benchmarks/visualize_benchmark_results.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import json
 
 import matplotlib.pyplot as plt

benchmarks/fused_kernels/layernorm_rms_benchmarks.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import pickle as pkl
 import time
 from dataclasses import dataclass

benchmarks/kernels/benchmark_aqlm.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import os
 import sys
 from typing import Optional

benchmarks/kernels/benchmark_layernorm.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import time
 
 import torch

benchmarks/kernels/benchmark_lora.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import copy
 import json

benchmarks/kernels/benchmark_machete.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import copy
 import itertools

benchmarks/kernels/benchmark_marlin.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from typing import List
 
 import torch

benchmarks/kernels/benchmark_moe.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import time
 from datetime import datetime

benchmarks/kernels/benchmark_paged_attention.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import random
 import time
 from typing import List, Optional

benchmarks/kernels/benchmark_quant.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import time
 
 import torch

benchmarks/kernels/benchmark_rmsnorm.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import itertools
 from typing import Optional, Tuple, Union
 

benchmarks/kernels/benchmark_rope.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from itertools import accumulate
 from typing import List, Optional
 

benchmarks/kernels/benchmark_shapes.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 WEIGHT_SHAPES = {
     "ideal": [[4 * 256 * 32, 256 * 32]],
     "mistralai/Mistral-7B-v0.1/TP1": [

benchmarks/kernels/graph_machete_bench.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import math
 import pickle
 import re

benchmarks/kernels/utils.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import dataclasses
 from typing import Any, Callable, Iterable, Optional
 

benchmarks/kernels/weight_shapes.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 # Weight Shapes are in the format
 # ([K, N], TP_SPLIT_DIM)
 # Example:

benchmarks/overheads/benchmark_hashing.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import cProfile
 import pstats
 

cmake/hipify.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 #!/usr/bin/env python3
 
 #

collect_env.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 # ruff: noqa
 # code borrowed from https://github.com/pytorch/pytorch/blob/main/torch/utils/collect_env.py
 

csrc/cutlass_extensions/vllm_cutlass_library_extension.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import enum
 from typing import Dict, Union
 

csrc/moe/moe_align_sum_kernels.cu

@@ -197,6 +197,72 @@ __global__ void moe_align_block_size_global_mem_kernel(
   }
 }
 
+// taken from
+// https://github.com/sgl-project/sglang/commit/ded9fcd09a43d5e7d5bb31a2bc3e9fc21bf65d2a
+template <typename scalar_t>
+__global__ void sgl_moe_align_block_size_kernel(
+    scalar_t* __restrict__ topk_ids, int32_t* sorted_token_ids,
+    int32_t* expert_ids, int32_t* total_tokens_post_pad, int32_t num_experts,
+    int32_t block_size, size_t numel, int32_t* cumsum) {
+  __shared__ int32_t shared_counts[32][8];
+  __shared__ int32_t local_offsets[256];
+
+  const int warp_id = threadIdx.x / WARP_SIZE;
+  const int lane_id = threadIdx.x % WARP_SIZE;
+  const int experts_per_warp = 8;
+  const int my_expert_start = warp_id * experts_per_warp;
+
+  for (int i = 0; i < experts_per_warp; ++i) {
+    if (my_expert_start + i < num_experts) {
+      shared_counts[warp_id][i] = 0;
+    }
+  }
+
+  const size_t tokens_per_thread = CEILDIV(numel, blockDim.x);
+  const size_t start_idx = threadIdx.x * tokens_per_thread;
+
+  for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+    int expert_id = topk_ids[i];
+    int warp_idx = expert_id / experts_per_warp;
+    int expert_offset = expert_id % experts_per_warp;
+    atomicAdd(&shared_counts[warp_idx][expert_offset], 1);
+  }
+
+  __syncthreads();
+
+  if (threadIdx.x == 0) {
+    cumsum[0] = 0;
+    for (int i = 1; i <= num_experts; ++i) {
+      int expert_count = 0;
+      int warp_idx = (i - 1) / experts_per_warp;
+      int expert_offset = (i - 1) % experts_per_warp;
+      expert_count = shared_counts[warp_idx][expert_offset];
+
+      cumsum[i] =
+          cumsum[i - 1] + CEILDIV(expert_count, block_size) * block_size;
+    }
+    *total_tokens_post_pad = cumsum[num_experts];
+  }
+
+  __syncthreads();
+
+  if (threadIdx.x < num_experts) {
+    for (int i = cumsum[threadIdx.x]; i < cumsum[threadIdx.x + 1];
+         i += block_size) {
+      expert_ids[i / block_size] = threadIdx.x;
+    }
+    local_offsets[threadIdx.x] = cumsum[threadIdx.x];
+  }
+
+  __syncthreads();
+
+  for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+    int32_t expert_id = topk_ids[i];
+    int32_t rank_post_pad = atomicAdd(&local_offsets[expert_id], 1);
+    sorted_token_ids[rank_post_pad] = i;
+  }
+}
+
 template <typename scalar_t, int TOPK>
 __global__ void moe_sum_kernel(
     scalar_t* __restrict__ out,          // [..., d]
@@ -305,6 +371,32 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
   }
 }
 
+void sgl_moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
+                              int64_t block_size,
+                              torch::Tensor sorted_token_ids,
+                              torch::Tensor experts_ids,
+                              torch::Tensor num_tokens_post_pad) {
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  VLLM_DISPATCH_INTEGRAL_TYPES(
+      topk_ids.scalar_type(), "sgl_moe_align_block_size_kernel", [&] {
+        // calc needed amount of shared mem for `tokens_cnts` and `cumsum`
+        // tensors
+        auto options_int =
+            torch::TensorOptions().dtype(torch::kInt).device(topk_ids.device());
+        // torch::Tensor token_cnts_buffer =
+        //     torch::empty({(num_experts + 1) * num_experts}, options_int);
+        torch::Tensor cumsum_buffer =
+            torch::empty({num_experts + 1}, options_int);
+
+        auto kernel = vllm::moe::sgl_moe_align_block_size_kernel<scalar_t>;
+        kernel<<<1, 1024, 0, stream>>>(
+            topk_ids.data_ptr<scalar_t>(), sorted_token_ids.data_ptr<int32_t>(),
+            experts_ids.data_ptr<int32_t>(),
+            num_tokens_post_pad.data_ptr<int32_t>(), num_experts, block_size,
+            topk_ids.numel(), cumsum_buffer.data_ptr<int32_t>());
+      });
+}
+
 void moe_sum(torch::Tensor& input,   // [num_tokens, topk, hidden_size]
              torch::Tensor& output)  // [num_tokens, hidden_size]
 {

csrc/moe/moe_ops.h

@@ -12,3 +12,9 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
                           int64_t block_size, torch::Tensor sorted_token_ids,
                           torch::Tensor experts_ids,
                           torch::Tensor num_tokens_post_pad);
+
+void sgl_moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
+                              int64_t block_size,
+                              torch::Tensor sorted_token_ids,
+                              torch::Tensor experts_ids,
+                              torch::Tensor num_tokens_post_pad);

csrc/moe/torch_bindings.cpp

@@ -22,6 +22,15 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
       "                     Tensor! num_tokens_post_pad) -> ()");
   m.impl("moe_align_block_size", torch::kCUDA, &moe_align_block_size);
 
+  // temporarily adapted from
+  // https://github.com/sgl-project/sglang/commit/ded9fcd09a43d5e7d5bb31a2bc3e9fc21bf65d2a
+  m.def(
+      "sgl_moe_align_block_size(Tensor topk_ids, int num_experts,"
+      "                         int block_size, Tensor! sorted_token_ids,"
+      "                         Tensor! experts_ids,"
+      "                         Tensor! num_tokens_post_pad) -> ()");
+  m.impl("sgl_moe_align_block_size", torch::kCUDA, &sgl_moe_align_block_size);
+
 #ifndef USE_ROCM
   m.def(
       "marlin_gemm_moe(Tensor! a, Tensor! b_q_weights, Tensor! sorted_ids, "

csrc/quantization/machete/generate.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import itertools
 import math
 import os

docs/source/conf.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 # Configuration file for the Sphinx documentation builder.
 #
 # This file only contains a selection of the most common options. For a full

docs/source/generate_examples.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import itertools
 import re
 from dataclasses import dataclass, field

docs/source/getting_started/installation/gpu/xpu.inc.md

@@ -36,7 +36,7 @@ VLLM_TARGET_DEVICE=xpu python setup.py install
 
 :::{note}
 - FP16 is the default data type in the current XPU backend. The BF16 data
-  type will be supported in the future.
+  type is supported on Intel Data Center GPU, not supported on Intel Arc GPU yet.
 :::
 
 ## Set up using Docker

docs/source/models/supported_models.md

@@ -40,6 +40,82 @@ If vLLM successfully returns text (for generative models) or hidden states (for
 Otherwise, please refer to [Adding a New Model](#new-model) for instructions on how to implement your model in vLLM.
 Alternatively, you can [open an issue on GitHub](https://github.com/vllm-project/vllm/issues/new/choose) to request vLLM support.
 
+### Transformers fallback
+
+After the merge of <gh-pr:11330>, `vllm` can fallback to models that are available in `transformers`. This does not work for all models for now, but most decoder language models are supported, and vision language model support is planned!
+
+To check if the backend is `transformers`, you can simply do this:
+
+```python 
+from vllm import LLM
+llm = LLM(model=..., task="generate")  # Name or path of your model
+llm.apply_model(lambda model: print(model.__class__))
+```
+
+If it is `TransformersModel` then it means it's based on `transformers`!
+
+#### Supported features
+
+##### LORA and quantization
+
+Both are not supported yet! Make sure to open an issue and we'll work on this together with the `transformers` team!
+
+Usually `transformers` model load weights via the `load_adapters` API, that depends on PEFT. We need to work a bit to either use this api (for now this would result in some weights not being marked as loaded) or replace modules accordingly.
+
+Hints as to how this would look like:
+
+```python
+class TransformersModel(nn.Module, SupportsLoRA):
+  def __init__(*):
+    ...
+    self.model.load_adapter(vllm_config.load_config.model_loader_extra_config["qlora_adapter_name_or_path"])
+```
+
+Blocker is that you need to specify supported lora layers, when we would ideally want to load whatever is inside the checkpoint!
+
+##### Remote code
+
+This fallback also means that any model on the hub that can be used in `transformers` with `trust_remote_code=True` that correctly implements attention can be used in production!
+
+```python 
+from vllm import LLM
+llm = LLM(model=..., task="generate", trust_remote_code=True)  # Name or path of your model
+llm.apply_model(lambda model: print(model.__class__))
+```
+
+A model just needs the following two things:
+
+```python
+from transformers import PreTrainedModel
+from torch import nn
+
+class MyAttention(nn.Module):
+
+  def forward(self, hidden_states, **kwargs): # <- kwargs are required
+
+    ...
+    attention_interface = attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+    attn_output, attn_weights = attention_interface(
+      self,
+      query_states,
+      key_states,
+      value_states,
+      **kwargs,
+    )
+    ...
+
+class MyModel(PreTrainedModel):
+  _supports_attention_backend = True
+```
+
+Here is what happens in the background:
+
+1. The config is loaded
+2. `MyModel` python class is loaded from the `auto_map`, and we check that the model `_supports_attention_backend`.
+3. The `TransformersModel` backend is used. See `/model_executors/models/transformers`, which leverage `self.config._attn_implementation = "vllm"`, thus the need to use `ALL_ATTENTION_FUNCTION`.
+
+That's it!
+
 ### ModelScope
 
 To use models from [ModelScope](https://www.modelscope.cn) instead of HuggingFace Hub, set an environment variable:

docs/source/serving/distributed_serving.md

@@ -60,7 +60,8 @@ bash run_cluster.sh \
                 vllm/vllm-openai \
                 ip_of_head_node \
                 --head \
-                /path/to/the/huggingface/home/in/this/node
+                /path/to/the/huggingface/home/in/this/node \
+                -e VLLM_HOST_IP=ip_of_this_node
 ```
 
 On the rest of the worker nodes, run the following command:
@@ -70,10 +71,11 @@ bash run_cluster.sh \
                 vllm/vllm-openai \
                 ip_of_head_node \
                 --worker \
-                /path/to/the/huggingface/home/in/this/node
+                /path/to/the/huggingface/home/in/this/node \
+                -e VLLM_HOST_IP=ip_of_this_node
 ```
 
-Then you get a ray cluster of containers. Note that you need to keep the shells running these commands alive to hold the cluster. Any shell disconnect will terminate the cluster. In addition, please note that the argument `ip_of_head_node` should be the IP address of the head node, which is accessible by all the worker nodes. A common misunderstanding is to use the IP address of the worker node, which is not correct.
+Then you get a ray cluster of containers. Note that you need to keep the shells running these commands alive to hold the cluster. Any shell disconnect will terminate the cluster. In addition, please note that the argument `ip_of_head_node` should be the IP address of the head node, which is accessible by all the worker nodes. The IP addresses of each worker node should be specified in the `VLLM_HOST_IP` environment variable, and should be different for each worker node. Please check the network configuration of your cluster to make sure the nodes can communicate with each other through the specified IP addresses.
 
 Then, on any node, use `docker exec -it node /bin/bash` to enter the container, execute `ray status` to check the status of the Ray cluster. You should see the right number of nodes and GPUs.
 
@@ -103,3 +105,7 @@ Please make sure you downloaded the model to all the nodes (with the same path),
 
 When you use huggingface repo id to refer to the model, you should append your huggingface token to the `run_cluster.sh` script, e.g. `-e HF_TOKEN=`. The recommended way is to download the model first, and then use the path to refer to the model.
 :::
+
+:::{warning}
+If you keep receiving the error message `Error: No available node types can fulfill resource request` but you have enough GPUs in the cluster, chances are your nodes have multiple IP addresses and vLLM cannot find the right one, especially when you are using multi-node inference. Please make sure vLLM and ray use the same IP address. You can set the `VLLM_HOST_IP` environment variable to the right IP address in the `run_cluster.sh` script (different for each node!), and check `ray status` to see the IP address used by Ray. See <gh-issue:7815> for more information.
+:::

examples/offline_inference/aqlm_example.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from vllm import LLM, SamplingParams
 from vllm.utils import FlexibleArgumentParser
 

examples/offline_inference/arctic.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from vllm import LLM, SamplingParams
 
 # Sample prompts.

examples/offline_inference/audio_language.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 This example shows how to use vLLM for running offline inference 
 with the correct prompt format on audio language models.

examples/offline_inference/basic.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from vllm import LLM, SamplingParams
 
 # Sample prompts.

examples/offline_inference/basic_with_model_default_sampling.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from vllm import LLM
 
 # Sample prompts.

examples/offline_inference/chat.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from vllm import LLM, SamplingParams
 
 llm = LLM(model="meta-llama/Meta-Llama-3-8B-Instruct")

examples/offline_inference/chat_with_tools.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 # ruff: noqa
 import json
 import random

examples/offline_inference/classification.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from vllm import LLM
 
 # Sample prompts.

examples/offline_inference/cli.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from dataclasses import asdict
 
 from vllm import LLM, SamplingParams

examples/offline_inference/cpu_offload.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from vllm import LLM, SamplingParams
 
 # Sample prompts.

examples/offline_inference/distributed.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 This example shows how to use Ray Data for running offline batch inference
 distributively on a multi-nodes cluster.

examples/offline_inference/embedding.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from vllm import LLM
 
 # Sample prompts.

examples/offline_inference/encoder_decoder.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 '''
 Demonstrate prompting of text-to-text
 encoder/decoder models, specifically BART

examples/offline_inference/florence2_inference.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 '''
 Demonstrate prompting of text-to-text
 encoder/decoder models, specifically Florence-2

examples/offline_inference/gguf_inference.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from huggingface_hub import hf_hub_download
 
 from vllm import LLM, SamplingParams

examples/offline_inference/llm_engine_example.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 from typing import List, Tuple
 

examples/offline_inference/lora_with_quantization_inference.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 This example shows how to use LoRA with different quantization techniques
 for offline inference.

examples/offline_inference/mlpspeculator.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import gc
 import time
 from typing import List

examples/offline_inference/multilora_inference.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 This example shows how to use the multi-LoRA functionality
 for offline inference.

examples/offline_inference/neuron.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from vllm import LLM, SamplingParams
 
 # Sample prompts.

examples/offline_inference/neuron_int8_quantization.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import os
 
 from vllm import LLM, SamplingParams

examples/offline_inference/pixtral.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 # ruff: noqa
 import argparse
 

examples/offline_inference/prefix_caching.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from vllm import LLM, SamplingParams
 from vllm.distributed import cleanup_dist_env_and_memory
 

examples/offline_inference/profiling.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import inspect
 import json
 import os

examples/offline_inference/profiling_tpu/profiling.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import dataclasses
 import os

examples/offline_inference/rlhf.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 a simple demonstration of RLHF with vLLM, inspired by
 the OpenRLHF framework https://github.com/OpenRLHF/OpenRLHF .

examples/offline_inference/save_sharded_state.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 Saves each worker's model state dict directly to a checkpoint, which enables a
 fast load path for large tensor-parallel models where each worker only needs to

examples/offline_inference/scoring.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from vllm import LLM
 
 # Sample prompts.

examples/offline_inference/simple_profiling.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import os
 import time
 

examples/offline_inference/structured_outputs.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from enum import Enum
 
 from pydantic import BaseModel

examples/offline_inference/torchrun_example.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 experimental support for tensor-parallel inference with torchrun,
 see https://github.com/vllm-project/vllm/issues/11400 for

examples/offline_inference/tpu.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from vllm import LLM, SamplingParams
 
 prompts = [

examples/offline_inference/vision_language.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 This example shows how to use vLLM for running offline inference with
 the correct prompt format on vision language models for text generation.

examples/offline_inference/vision_language_embedding.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 This example shows how to use vLLM for running offline inference with
 the correct prompt format on vision language models for multimodal embedding.

examples/offline_inference/vision_language_multi_image.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 This example shows how to use vLLM for running offline inference with
 multi-image input on vision language models for text generation,

examples/offline_inference/whisper.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import time
 
 from vllm import LLM, SamplingParams

examples/online_serving/api_client.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """Example Python client for `vllm.entrypoints.api_server`
 NOTE: The API server is used only for demonstration and simple performance
 benchmarks. It is not intended for production use.

examples/online_serving/cohere_rerank_client.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 Example of using the OpenAI entrypoint's rerank API which is compatible with
 the Cohere SDK: https://github.com/cohere-ai/cohere-python

examples/online_serving/gradio_openai_chatbot_webserver.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 
 import gradio as gr

examples/online_serving/gradio_webserver.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import json
 

examples/online_serving/jinaai_rerank_client.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 Example of using the OpenAI entrypoint's rerank API which is compatible with
 Jina and Cohere https://jina.ai/reranker

examples/online_serving/openai_chat_completion_client.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from openai import OpenAI
 
 # Modify OpenAI's API key and API base to use vLLM's API server.