Skip to content

Commit

Permalink
[Documentation][Spec Decode] Add documentation about lossless guarant…
Browse files Browse the repository at this point in the history
…ees in Speculative Decoding in vLLM (vllm-project#7962)
  • Loading branch information
sroy745 authored Sep 5, 2024
1 parent 2ee4528 commit 2febcf2
Show file tree
Hide file tree
Showing 2 changed files with 59 additions and 0 deletions.
40 changes: 40 additions & 0 deletions docs/source/models/spec_decode.rst
Original file line number Diff line number Diff line change
Expand Up @@ -161,6 +161,46 @@ A variety of speculative models of this type are available on HF hub:
* `granite-7b-instruct-accelerator <https://huggingface.co/ibm-granite/granite-7b-instruct-accelerator>`_
* `granite-20b-code-instruct-accelerator <https://huggingface.co/ibm-granite/granite-20b-code-instruct-accelerator>`_

Lossless guarantees of Speculative Decoding
-------------------------------------------
In vLLM, speculative decoding aims to enhance inference efficiency while maintaining accuracy. This section addresses the lossless guarantees of
speculative decoding, breaking down the guarantees into three key areas:

1. **Theoretical Losslessness**
- Speculative decoding sampling is theoretically lossless up to the precision limits of hardware numerics. Floating-point errors might
cause slight variations in output distributions, as discussed
in `Accelerating Large Language Model Decoding with Speculative Sampling <https://arxiv.org/pdf/2302.01318>`_

2. **Algorithmic Losslessness**
- vLLM’s implementation of speculative decoding is algorithmically validated to be lossless. Key validation tests include:

- **Rejection Sampler Convergence**: Ensures that samples from vLLM’s rejection sampler align with the target
distribution. `View Test Code <https://github.com/vllm-project/vllm/blob/47b65a550866c7ffbd076ecb74106714838ce7da/tests/samplers/test_rejection_sampler.py#L252>`_

- **Greedy Sampling Equality**: Confirms that greedy sampling with speculative decoding matches greedy sampling
without it. This verifies that vLLM's speculative decoding framework, when integrated with the vLLM forward pass and the vLLM rejection sampler,
provides a lossless guarantee. Almost all of the tests in `this directory <https://github.com/vllm-project/vllm/tree/b67ae00cdbbe1a58ffc8ff170f0c8d79044a684a/tests/spec_decode/e2e>`_
verify this property using `this assertion implementation <https://github.com/vllm-project/vllm/blob/b67ae00cdbbe1a58ffc8ff170f0c8d79044a684a/tests/spec_decode/e2e/conftest.py#L291>`_

3. **vLLM Logprob Stability**
- vLLM does not currently guarantee stable token log probabilities (logprobs). This can result in different outputs for the
same request across runs. For more details, see the FAQ section
titled *Can the output of a prompt vary across runs in vLLM?* in the `FAQs <../serving/faq.rst>`_.


**Conclusion**

While vLLM strives to ensure losslessness in speculative decoding, variations in generated outputs with and without speculative decoding
can occur due to following factors:

- **Floating-Point Precision**: Differences in hardware numerical precision may lead to slight discrepancies in the output distribution.

- **Batch Size and Numerical Stability**: Changes in batch size may cause variations in logprobs and output probabilities, potentially
due to non-deterministic behavior in batched operations or numerical instability.

**Mitigation Strategies**

For mitigation strategies, please refer to the FAQ entry *Can the output of a prompt vary across runs in vLLM?* in the `FAQs <../serving/faq.rst>`_.

Resources for vLLM contributors
-------------------------------
Expand Down
19 changes: 19 additions & 0 deletions docs/source/serving/faq.rst
Original file line number Diff line number Diff line change
Expand Up @@ -10,3 +10,22 @@ A: Assuming that you're referring to using OpenAI compatible server to serve mul
Q: Which model to use for offline inference embedding?

A: If you want to use an embedding model, try: https://huggingface.co/intfloat/e5-mistral-7b-instruct. Instead models, such as Llama-3-8b, Mistral-7B-Instruct-v0.3, are generation models rather than an embedding model

----------------------------------------

Q: Can the output of a prompt vary across runs in vLLM?

A: Yes, it can. vLLM does not guarantee stable log probabilities (logprobs) for the output tokens. Variations in logprobs may occur due to
numerical instability in Torch operations or non-deterministic behavior in batched Torch operations when batching changes. For more details,
see the `Numerical Accuracy section <https://pytorch.org/docs/stable/notes/numerical_accuracy.html#batched-computations-or-slice-computations>`_.

In vLLM, the same requests might be batched differently due to factors such as other concurrent requests,
changes in batch size, or batch expansion in speculative decoding. These batching variations, combined with numerical instability of Torch operations,
can lead to slightly different logit/logprob values at each step. Such differences can accumulate, potentially resulting in
different tokens being sampled. Once a different token is sampled, further divergence is likely.

**Mitigation Strategies**

- For improved stability and reduced variance, use `float32`. Note that this will require more memory.
- If using `bfloat16`, switching to `float16` can also help.
- Using request seeds can aid in achieving more stable generation for temperature > 0, but discrepancies due to precision differences may still occur.

0 comments on commit 2febcf2

Please sign in to comment.