Llama 2

Language: Mojo 🔥

API: MAX Graph

This pipeline demonstrates text completion from an initial prompt using the Llama 2 large language model. The model itself has been constructed from end to end in the Mojo language using the MAX Graph API.

The MAX Graph API provides an accessible Mojo interface to the construction of flexible accelerated compute graphs, which are then optimized by the MAX Engine's advanced graph compiler. This pipeline showcases how a large language model can be fully defined using Mojo and MAX Graphs and then compiled for optimal inference performance via the MAX Engine.

Model

Llama 2 is an open source large language model released by Meta. The structure of this implementation was inspired by Andrej Karpathy's llama2.c, and originally written in Mojo by Aydyn Tairov.

The text completion demo is compatible with the the official Llama 2 text completion demo.

The default settings for this pipeline use the 7B set of pretrained weights in q4_k quantized encodings.

Usage

The easiest way to try out this pipeline is with our Magic command-line tool. Follow the instructions to install Magic. Once installed, you can try out text generation using Llama 2 with the following command:

magic run llama2 --prompt "I believe the meaning of life is"

On first execution, the tokenizer library and model weights will be downloaded and placed in a .cache/modular subdirectory within your home directory. The model will then be compiled and text completion will begin from the specified prompt.

To modify or build upon the pipeline code, you can use the following steps:

Install MAX:

If MAX is not already installed, follow the installation instructions to set it up on your system.
Clone the MAX examples repository:

If you don't already have a local clone of this repository, create one via:
```
git clone https://github.com/modularml/max.git
```
The following instructions assume that you're present within this pipeline's directory, and you can change to it after cloning:
```
cd max/examples/graph-api/pipelines/llama2/
```
(Optional) Install Python dependencies:

This enables using the HuggingFace transformers AutoTokenizer. If transformers isn't found, a Mojo tokenizer implementation is used.
```
python3 -m pip install -r requirements.txt
```
Run the text completion demo:

To access the llama models, you need to agree to their license in Huggingface.

License is located here Llama-2-7b-hf

All of the pipelines have been configured to use a common driver, located in the directory hosting all MAX Graph examples. Assuming you're starting at the path of this README, the command invocation will look like:
```
mojo ../../run_pipeline.🔥 llama2 --prompt "I believe the meaning of life is"
```

Options

The following command-line options are available to customize operation of the pipeline:

--model-path: Overrides the default URL, and allows for an already-downloaded pretrained weight file to be used with the model.
--custom-ops-path: The path to a compiled Mojo package containing a custom graph operation to use within the pipeline.
--tokenizer-path: The path to the tokenizer library to be used by the pipeline. (Default value: .cache/tokenizer.bin)
--max-length: The context length of the model. (Default value: 512)
--max-new-tokens: The maximum number of new tokens to generate. If a -1 value is provided, the model will continue to generate tokens for the entire context length. (Default value: -1)
--min-p: The starting required percentage for Min P sampling. (Default value: 0.05)
--prompt: The text prompt to use for further generation.
--quantization-encoding: The encoding to use for a datatype that can be quantized to a low bits per weight format. The options for quantized formats will download and cache default weights, but float32 requires the use of --model-path to specify locally downloaded full-precision weights for use in the model. Valid values: q4_0, q4_k, q6_k, float32. (Default value: q4_0).
--temperature: The temperature for sampling, on a scale from 0.0 - 1.0, with 0.0 being greedy sampling. (Default value: 0.5)

Ideas for future extension

There are many ways that this pipeline can be built upon or extended, and this is a short list of suggestions for future work:

Enhance the tokenizer so that it can stand alone as a general-purpose tokenizer for multiple text generation pipelines.
Expand the customizable options for text generation.
Incorporate and use weights from other models.
Improve the quality of the text generation.
Identify performance bottlenecks and further tune time-to-first-token and throughput.

501 lines - 15.68 KB

Mojo

Graph

# ===----------------------------------------------------------------------=== #
# Copyright (c) 2024, Modular Inc. All rights reserved.
#
# Licensed under the Apache License v2.0 with LLVM Exceptions:
# https://llvm.org/LICENSE.txt
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ===----------------------------------------------------------------------=== #

from collections import Dict
import sys
from math import align_up
from memory import memcpy
from os import setenv
from pathlib import Path, cwd
from utils.index import Index

from max.engine import InferenceSession, Model, SessionOptions
from max.driver import (
    AnyTensor,
    AnyMojoValue,
    Device,
    Tensor,
    cpu_device,
)
from max.tensor import TensorShape
from max.driver._cuda import cuda_device
from max.graph import Graph
from max.graph.quantization import (
    Float32Encoding,
    QuantizationEncoding,
    Q4_0Encoding,
    Q4_KEncoding,
    Q6_KEncoding,
)

from max.serve.kv_cache.types import (
    ContiguousKVCache,
    ContiguousKVCacheManager,
    ContiguousKVCacheCollection,
    KVCacheLayout,
    KVCacheStaticParams,
)

from ..llama3.kv_cache import KVCache
from ..llama3.metrics import Metrics
from .model import Llama2, QuantizedLlama2
from ..samplers.weighted_sampler import WeightedSampler
from .tokenizer.bpe import BPETokenizer
from ..configs.common import get_max_tokens_to_generate
from ..configs.llama import (
    get_llama2_model_url,
    LlamaConfigRegistry,
    get_llama_base_default_config,
)
from ..configs.registry import ConfigRegistryDict
from ..configs.parse_args import (
    OptionTypeEnum,
    OptionValue,
    parse_args,
    register_pipeline_configs,
)
from ..tokenizer import AutoTokenizer, Tokenizer
from ..weights.download import download_to_cache, modular_cache_dir
from ..weights.gguf import GGUFFile
from ..weights.llama2checkpoint import LlamaCFile
from ..weights.loadable_model import LlamaHParams, LoadableModel


@value
struct Config:
    """Configuration for token generation runtime options."""

    var config: Dict[String, OptionValue]

    def __init__(inout self):
        config_registry = LlamaConfigRegistry(ConfigRegistryDict())

        default_configs = get_llama_base_default_config()
        self.config = register_pipeline_configs(
            config_registry.registry,
            parse_args(),
            default_configs,
        )

        # Check for invalid config
        model_path = self.config["model-path"]
        quantization_encoding = self.config["quantization-encoding"]
        if (
            model_path[Path].suffix() != ".gguf"
            and not quantization_encoding[String]
        ):
            raise (
                "`--model-path` must be `.bin` or `.gguf` file. "
                "Alternatively provide a `--quantization-encoding`"
            )

    fn get(inout self, key: String) raises -> OptionValue:
        """Returns an option value for `key` in the underlying config.

        Args:
            key: Key for the underlying config option.

        Returns:
            An OptionValue.

        Raises:
            An error for invalid key.
        """
        return self.config[key]

    fn set(inout self, key: String, val: OptionValue):
        """Sets a new value for a given config key. This will overwrite the old
        value if the key is already present.

        Args:
            key: A string based key for the underlying config option.
            val: A new value for a key that already exist.
        """
        self.config[key] = val


def compile_graph(
    graph: Graph,
    execution_device: Device,
    custom_ops_paths: List[Path] = List[Path](),
) -> Model:
    """Compiles a staged graph using the graph compiler."""
    session = InferenceSession(SessionOptions(execution_device))
    print("Compiling...")
    return session.load(graph, custom_ops_paths=custom_ops_paths)


def _get_attention_mask(
    prompt_mask: Tensor[DType.bool, 2], n: Int, host_device: Device
) -> AnyTensor:
    mask = Tensor[DType.bool, rank=2](TensorShape(1, n), host_device)
    memcpy(mask.unsafe_ptr(), prompt_mask.unsafe_ptr(), prompt_mask.spec()[1])
    for i in range(prompt_mask.spec()[1], n):
        mask[0, i] = True
    return mask


def _generate_q_text_with_tokenizer[
    tokenizer_type: Tokenizer,
](
    inout tokenizer: tokenizer_type,
    compiled_model: Model,
    params: LlamaHParams,
    config: Config,
    inout metrics: Metrics,
    execution_device: Device,
):
    host_device = cpu_device()

    metrics.begin_timing_prompt()
    prompt = tokenizer.encode(
        config.get("prompt")[String], bos=String("\n<s>\n")
    )
    padded_size = align_up(prompt.size, config.get("pad-to-multiple-of")[Int])
    n_pad_tokens = padded_size - prompt.size
    metrics.set_tokens_in_prompt(padded_size)
    sampler = WeightedSampler(
        config.get("temperature")[Float64].cast[DType.float32](),
        config.get("min-p")[Float64].cast[DType.float32](),
    )

    # Allocate input & attention mask tensors, then initialize them.
    # FIXME (MSDK-774): Padding logic should be handled by tokenizer instead.
    tokens = Tensor[DType.int64, rank=2](
        TensorShape(1, padded_size), host_device
    )
    prompt_attn_mask = Tensor[DType.bool, rank=2](
        TensorShape(1, padded_size), host_device
    )
    for i in range(padded_size):
        tokens[0, i] = 0 if i < n_pad_tokens else prompt[i - n_pad_tokens]
        prompt_attn_mask[0, i] = False if i < n_pad_tokens else True

    print("Executing...")
    print(tokenizer.decode(prompt), end="")

    kv_cache = KVCache(
        params,
        config.get("max-length")[Int],
        config.get("batch-size")[Int],
        host_device,
    )

    # The first iteration caches the entire prompt and all subsequent
    # iterations generate one token.
    # Avoid overrunning the cache by setting the trip count accordingly.
    metrics.begin_timing_generation()
    max_tokens = get_max_tokens_to_generate(
        padded_size,
        config.get("max-length")[Int],
        config.get("max-new-tokens")[Int],
    )
    for i in range(padded_size, max_tokens + 1):
        results = compiled_model.execute(
            tokens.to_device_tensor().move_to(execution_device),
            _get_attention_mask(prompt_attn_mask, i, host_device)
            .to_device_tensor()
            .move_to(execution_device),
            kv_cache.keys_view(execution_device),
            kv_cache.values_view(execution_device),
        )

        kv_cache.update(results[1].take(), results[2].take())

        logits = results[0].take().to_device_tensor()
        logits = logits.move_to(host_device)
        logits_tensor = logits.to_tensor[DType.float32, rank=2]()
        token = Int64(sampler.sample(logits_tensor^).selected)

        tokens = Tensor[DType.int64, rank=2](TensorShape(1, 1), host_device)
        tokens[0, 0] = token

        metrics.new_token()
        print(tokenizer.decode(token), end="")
    print()
    metrics.end_timing()


def _generate_text_with_tokenizer[
    tokenizer_type: Tokenizer,
    kv_params: KVCacheStaticParams,
](
    inout tokenizer: tokenizer_type,
    compiled_model: Model,
    params: LlamaHParams,
    config: Config,
    inout metrics: Metrics,
    execution_device: Device,
):
    host_device = cpu_device()
    metrics.begin_timing_prompt()

    # Encode prompt and left pad-to-multiple-of
    prompt = tokenizer.encode(
        config.get("prompt")[String], bos=String("\n<s>\n")
    )
    padded_size = align_up(prompt.size, config.get("pad-to-multiple-of")[Int])
    n_pad_tokens = padded_size - prompt.size

    metrics.set_tokens_in_prompt(padded_size)

    # Allocate input & attention mask tensors, then initialize them.
    # FIXME (MSDK-774): Padding logic should be handled by tokenizer instead.
    tokens = Tensor[DType.int64, rank=2](
        TensorShape(1, padded_size), host_device
    )
    prompt_attn_mask = Tensor[DType.bool, rank=2](
        TensorShape(1, padded_size), host_device
    )
    for i in range(padded_size):
        tokens[0, i] = 0 if i < n_pad_tokens else prompt[i - n_pad_tokens]
        prompt_attn_mask[0, i] = False if i < n_pad_tokens else True

    max_tokens = get_max_tokens_to_generate(
        padded_size,
        config.get("max-length")[Int],
        config.get("max-new-tokens")[Int],
    )

    # Initialize Sampler
    sampler = WeightedSampler(
        config.get("temperature")[Float64].cast[DType.float32](),
        config.get("min-p")[Float64].cast[DType.float32](),
    )

    print("Executing...")
    print(tokenizer.decode(prompt), end="")

    kv_manager = ContiguousKVCacheManager[DType.float32, kv_params,](
        config.get("batch-size")[Int],
        max_tokens,
        params.n_layers,
        execution_device,
        host_device,
    )

    kv_collection = kv_manager.claim(config.get("batch-size")[Int])

    # The first iteration caches the entire prompt and all subsequent
    # iterations generate one token.
    # Avoid overrunning the cache by setting the trip count accordingly.
    metrics.begin_timing_generation()
    for i in range(padded_size, max_tokens + 1):
        if i == padded_size:
            valid_lengths = List[Int](padded_size)
        else:
            valid_lengths = List[Int](1)

        # Update Mask
        mask = Tensor[DType.bool, rank=2]((1, i), host_device)
        memcpy(
            mask.unsafe_ptr(),
            prompt_attn_mask.unsafe_ptr(),
            prompt_attn_mask.spec()[1],
        )
        for j in range(prompt_attn_mask.spec()[1], i):
            mask[0, j] = True

        result = compiled_model.execute(
            tokens.move_to(execution_device),
            mask.move_to(execution_device),
            AnyMojoValue(kv_collection^),
        )

        logits = (
            result[0]
            .take()
            .to_device_tensor()
            .move_to(host_device)
            .to_tensor[DType.float32, 2]()
        )

        kv_collection = (
            result[1]
            .take()
            .to[ContiguousKVCacheCollection[DType.float32, kv_params]]()
        )

        kv_manager.step(valid_lengths, kv_collection)

        token = SIMD[DType.int64, 1](sampler.sample(logits).selected)
        tokens = Tensor[DType.int64, rank=2]((1, 1), host_device)
        tokens[0, 0] = token

        metrics.new_token()
        print(tokenizer.decode(token), end="")
    print()
    metrics.end_timing()

    _ = kv_manager^
    _ = sampler^


# TODO: Delete this when we clean up quantize-tinystories
def generate_text[
    layout: KVCacheLayout
](
    compiled_model: Model,
    params: LlamaHParams,
    config: Config,
    execution_device: Device,
    inout metrics: Metrics,
):
    """Generated text by applying the compiled model to the provided prompt."""
    mojo_tokenizer = BPETokenizer.from_file(config.get("tokenizer-path")[Path])
    if params.n_kv_heads == 6 and params.head_dim == 48:
        _generate_text_with_tokenizer[
            BPETokenizer,
            KVCacheStaticParams(num_heads=6, head_size=48, layout=layout),
        ](
            mojo_tokenizer,
            compiled_model,
            params,
            config=config,
            execution_device=execution_device,
            metrics=metrics,
        )
    else:
        raise "Unsupported n_kv_head (" + str(
            params.n_kv_heads
        ) + ") and head_dim (" + str(params.head_dim) + ")"


def run[
    model_type: LoadableModel,
    encoding: QuantizationEncoding,
    target: StringLiteral,
    kv_params: KVCacheStaticParams,
](config: Config) -> None:
    # Initialize Device
    execution_device = cpu_device() if target == "cpu" else cuda_device()

    print("Building model...")
    metrics = Metrics()
    metrics.begin_timing_startup()

    model_params = model_type(config.get("model-path")[Path])
    params = model_params.hyperparams()

    model = Llama2[model_type, kv_params, encoding](
        model_params^,
    )
    graph = model.build_graph("llama2")

    compiled_model = compile_graph(
        graph, execution_device, config.get("custom-ops-path")[List[Path]]
    )
    metrics.end_timing_startup()

    # Get Tokenizer
    mojo_tokenizer = BPETokenizer.from_file(config.get("tokenizer-path")[Path])
    _generate_text_with_tokenizer[BPETokenizer, kv_params](
        mojo_tokenizer,
        compiled_model,
        params,
        config=config,
        execution_device=execution_device,
        metrics=metrics,
    )


def runq[
    model_type: LoadableModel, encoding: QuantizationEncoding
](config: Config) -> None:
    # initialize Metrics
    metrics = Metrics()
    metrics.begin_timing_startup()

    # Build Model
    model = QuantizedLlama2[model_type, encoding](
        config.get("model-path")[Path],
    )
    params = model.hyperparams()
    graph = model.build_graph("llama2")

    # Quantized Llama can only be run on CPU
    execution_device = cpu_device()
    session_options = SessionOptions(execution_device)
    session = InferenceSession(session_options)

    compiled_model = compile_graph(
        graph, execution_device, config.get("custom-ops-path")[List[Path]]
    )

    metrics.end_timing_startup()

    mojo_tokenizer = BPETokenizer.from_file(config.get("tokenizer-path")[Path])
    _generate_q_text_with_tokenizer[BPETokenizer](
        mojo_tokenizer,
        compiled_model,
        params,
        config=config,
        execution_device=execution_device,
        metrics=metrics,
    )


def llama2_run():
    config = Config()
    encoding = config.get("quantization-encoding")[String]

    # Download Model and Tokenizer as Needed
    if not config.get("model-path")[Path]:
        model_path = download_to_cache(get_llama2_model_url(encoding))
        config.set("model-path", model_path)

    if not config.get("tokenizer-path")[Path]:
        tokenizer_path = download_to_cache(
            "https://github.com/tairov/llama2.mojo/raw/master/tokenizer.bin",
        )
        config.set("tokenizer-path", tokenizer_path)

    # Print CLI Warnings
    if config.get("prompt")[String] == "I believe the meaning of life is":
        print("Using default prompt, provide an argument to change it:")
        print('    --prompt "Hello llama3"')

    # Fork Pipeline Runs
    if encoding != Float32Encoding.id():
        if config.get("experimental-use-gpu")[Bool]:
            raise encoding + " not available with 'experimental-use-gpu' option."

        if encoding == Q4_0Encoding.id():
            runq[GGUFFile, Q4_0Encoding](config)
        elif encoding == Q4_KEncoding.id():
            runq[GGUFFile, Q4_KEncoding](config)
        elif encoding == Q6_KEncoding.id():
            runq[GGUFFile, Q6_KEncoding](config)
        else:
            raise "--quantization-encoding " + encoding + " not supported"

    else:
        if config.get("experimental-use-gpu")[Bool]:
            run[
                LlamaCFile,
                Float32Encoding,
                "cuda",
                KVCacheStaticParams(
                    num_heads=6, head_size=48, layout=KVCacheLayout.BSHD
                ),
            ](config)
        else:
            run[
                LlamaCFile,
                Float32Encoding,
                "cpu",
                KVCacheStaticParams(
                    num_heads=6, head_size=48, layout=KVCacheLayout.BHSD
                ),
            ](config)

============================================================================================== The MAX repository is licensed under the Apache License v2.0 with LLVM Exceptions:

                             Apache License
                       Version 2.0, January 2004
                    http://www.apache.org/licenses/

TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION

1. Definitions.

  "License" shall mean the terms and conditions for use, reproduction,
  and distribution as defined by Sections 1 through 9 of this document.

  "Licensor" shall mean the copyright owner or entity authorized by
  the copyright owner that is granting the License.

  "Legal Entity" shall mean the union of the acting entity and all
  other entities that control, are controlled by, or are under common
  control with that entity. For the purposes of this definition,
  "control" means (i) the power, direct or indirect, to cause the
  direction or management of such entity, whether by contract or
  otherwise, or (ii) ownership of fifty percent (50%) or more of the
  outstanding shares, or (iii) beneficial ownership of such entity.

  "You" (or "Your") shall mean an individual or Legal Entity
  exercising permissions granted by this License.

  "Source" form shall mean the preferred form for making modifications,
  including but not limited to software source code, documentation
  source, and configuration files.

  "Object" form shall mean any form resulting from mechanical
  transformation or translation of a Source form, including but
  not limited to compiled object code, generated documentation,
  and conversions to other media types.

  "Work" shall mean the work of authorship, whether in Source or
  Object form, made available under the License, as indicated by a
  copyright notice that is included in or attached to the work
  (an example is provided in the Appendix below).

  "Derivative Works" shall mean any work, whether in Source or Object
  form, that is based on (or derived from) the Work and for which the
  editorial revisions, annotations, elaborations, or other modifications
  represent, as a whole, an original work of authorship. For the purposes
  of this License, Derivative Works shall not include works that remain
  separable from, or merely link (or bind by name) to the interfaces of,
  the Work and Derivative Works thereof.

  "Contribution" shall mean any work of authorship, including
  the original version of the Work and any modifications or additions
  to that Work or Derivative Works thereof, that is intentionally
  submitted to Licensor for inclusion in the Work by the copyright owner
  or by an individual or Legal Entity authorized to submit on behalf of
  the copyright owner. For the purposes of this definition, "submitted"
  means any form of electronic, verbal, or written communication sent
  to the Licensor or its representatives, including but not limited to
  communication on electronic mailing lists, source code control systems,
  and issue tracking systems that are managed by, or on behalf of, the
  Licensor for the purpose of discussing and improving the Work, but
  excluding communication that is conspicuously marked or otherwise
  designated in writing by the copyright owner as "Not a Contribution."

  "Contributor" shall mean Licensor and any individual or Legal Entity
  on behalf of whom a Contribution has been received by Licensor and
  subsequently incorporated within the Work.

2. Grant of Copyright License. Subject to the terms and conditions of
  this License, each Contributor hereby grants to You a perpetual,
  worldwide, non-exclusive, no-charge, royalty-free, irrevocable
  copyright license to reproduce, prepare Derivative Works of,
  publicly display, publicly perform, sublicense, and distribute the
  Work and such Derivative Works in Source or Object form.

3. Grant of Patent License. Subject to the terms and conditions of
  this License, each Contributor hereby grants to You a perpetual,
  worldwide, non-exclusive, no-charge, royalty-free, irrevocable
  (except as stated in this section) patent license to make, have made,
  use, offer to sell, sell, import, and otherwise transfer the Work,
  where such license applies only to those patent claims licensable
  by such Contributor that are necessarily infringed by their
  Contribution(s) alone or by combination of their Contribution(s)
  with the Work to which such Contribution(s) was submitted. If You
  institute patent litigation against any entity (including a
  cross-claim or counterclaim in a lawsuit) alleging that the Work
  or a Contribution incorporated within the Work constitutes direct
  or contributory patent infringement, then any patent licenses
  granted to You under this License for that Work shall terminate
  as of the date such litigation is filed.

4. Redistribution. You may reproduce and distribute copies of the
  Work or Derivative Works thereof in any medium, with or without
  modifications, and in Source or Object form, provided that You
  meet the following conditions:

  (a) You must give any other recipients of the Work or
      Derivative Works a copy of this License; and

  (b) You must cause any modified files to carry prominent notices
      stating that You changed the files; and

  (c) You must retain, in the Source form of any Derivative Works
      that You distribute, all copyright, patent, trademark, and
      attribution notices from the Source form of the Work,
      excluding those notices that do not pertain to any part of
      the Derivative Works; and

  (d) If the Work includes a "NOTICE" text file as part of its
      distribution, then any Derivative Works that You distribute must
      include a readable copy of the attribution notices contained
      within such NOTICE file, excluding those notices that do not
      pertain to any part of the Derivative Works, in at least one
      of the following places: within a NOTICE text file distributed
      as part of the Derivative Works; within the Source form or
      documentation, if provided along with the Derivative Works; or,
      within a display generated by the Derivative Works, if and
      wherever such third-party notices normally appear. The contents
      of the NOTICE file are for informational purposes only and
      do not modify the License. You may add Your own attribution
      notices within Derivative Works that You distribute, alongside
      or as an addendum to the NOTICE text from the Work, provided
      that such additional attribution notices cannot be construed
      as modifying the License.

  You may add Your own copyright statement to Your modifications and
  may provide additional or different license terms and conditions
  for use, reproduction, or distribution of Your modifications, or
  for any such Derivative Works as a whole, provided Your use,
  reproduction, and distribution of the Work otherwise complies with
  the conditions stated in this License.

5. Submission of Contributions. Unless You explicitly state otherwise,
  any Contribution intentionally submitted for inclusion in the Work
  by You to the Licensor shall be under the terms and conditions of
  this License, without any additional terms or conditions.
  Notwithstanding the above, nothing herein shall supersede or modify
  the terms of any separate license agreement you may have executed
  with Licensor regarding such Contributions.

6. Trademarks. This License does not grant permission to use the trade
  names, trademarks, service marks, or product names of the Licensor,
  except as required for reasonable and customary use in describing the
  origin of the Work and reproducing the content of the NOTICE file.

7. Disclaimer of Warranty. Unless required by applicable law or
  agreed to in writing, Licensor provides the Work (and each
  Contributor provides its Contributions) on an "AS IS" BASIS,
  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
  implied, including, without limitation, any warranties or conditions
  of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
  PARTICULAR PURPOSE. You are solely responsible for determining the
  appropriateness of using or redistributing the Work and assume any
  risks associated with Your exercise of permissions under this License.

8. Limitation of Liability. In no event and under no legal theory,
  whether in tort (including negligence), contract, or otherwise,
  unless required by applicable law (such as deliberate and grossly
  negligent acts) or agreed to in writing, shall any Contributor be
  liable to You for damages, including any direct, indirect, special,
  incidental, or consequential damages of any character arising as a
  result of this License or out of the use or inability to use the
  Work (including but not limited to damages for loss of goodwill,
  work stoppage, computer failure or malfunction, or any and all
  other commercial damages or losses), even if such Contributor
  has been advised of the possibility of such damages.

9. Accepting Warranty or Additional Liability. While redistributing
  the Work or Derivative Works thereof, You may choose to offer,
  and charge a fee for, acceptance of support, warranty, indemnity,
  or other liability obligations and/or rights consistent with this
  License. However, in accepting such obligations, You may act only
  on Your own behalf and on Your sole responsibility, not on behalf
  of any other Contributor, and only if You agree to indemnify,
  defend, and hold each Contributor harmless for any liability
  incurred by, or claims asserted against, such Contributor by reason
  of your accepting any such warranty or additional liability.

END OF TERMS AND CONDITIONS

APPENDIX: How to apply the Apache License to your work.

  To apply the Apache License to your work, attach the following
  boilerplate notice, with the fields enclosed by brackets "[]"
  replaced with your own identifying information. (Don't include
  the brackets!)  The text should be enclosed in the appropriate
  comment syntax for the file format. We also recommend that a
  file or class name and description of purpose be included on the
  same "printed page" as the copyright notice for easier
  identification within third-party archives.

Copyright [yyyy] [name of copyright owner]

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

   http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

---- LLVM Exceptions to the Apache 2.0 License ----

As an exception, if, as a result of your compiling your source code, portions of this Software are embedded into an Object form of such source code, you may redistribute such embedded portions in such Object form without complying with the conditions of Sections 4(a), 4(b) and 4(d) of the License.

In addition, if you combine or link compiled forms of this Software with software that is licensed under the GPLv2 ("Combined Software") and if a court of competent jurisdiction determines that the patent provision (Section 3), the indemnity provision (Section 9) or other Section of the License conflicts with the conditions of the GPLv2, you may retroactively and prospectively choose to deem waived or otherwise exclude such Section(s) of the License, but only in their entirety and only with respect to the Combined Software.

============================================================================== Software from third parties included in the LLVM Project:

The LLVM Project contains third party software which is under different license terms. All such code will be identified clearly using at least one of two mechanisms:

It will be in a separate directory tree with its own LICENSE.txt or LICENSE file at the top containing the specific license and restrictions which apply to that software, or
It will contain specific license and restriction terms at the top of every file.