[KV Cache] support kv cache int8 per channel quantization #1663
Conversation
Signed-off-by: evian <eviantai@u.nus.edu>
|
👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review. Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed. |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @Eviannn, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!
This pull request introduces comprehensive support for int8 per-channel quantization of Key-Value (KV) cache tensors. This involves strategically reshaping the KV cache tensors before quantization to allow for per-channel scaling across the combined head and dimension axis, and then reshaping them back to their original form. Additionally, the quantization observer logic has been enhanced to correctly compute quantization parameters for these newly shaped tensors, ensuring accurate and efficient quantization for LLMs.
Highlights
- KV Cache Per-Channel Quantization Support: Implemented support for int8 per-channel quantization of Key-Value (KV) cache tensors, enabling more granular quantization for improved model efficiency and reduced memory footprint.
- Tensor Reshaping for Per-Channel Quantization: Introduced pre-quantization reshaping of KV cache tensors from
[batch, num_heads, seq_len, head_dim]to[batch, seq_len, num_heads * head_dim]and subsequent reshaping back after de-quantization. This transformation facilitates per-channel scaling across the combined head and dimension axis. - Dynamic Quantization Parameter Calculation: Updated the quantization observer to dynamically determine the correct dimension for calculating per-channel quantization parameters. This enhancement specifically accommodates the new 3D shape of KV cache tensors, ensuring accurate scale and zero-point computation.
Using Gemini Code Assist
The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.
Invoking Gemini
You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.
| Feature | Command | Description |
|---|---|---|
| Code Review | /gemini review |
Performs a code review for the current pull request in its current state. |
| Pull Request Summary | /gemini summary |
Provides a summary of the current pull request in its current state. |
| Comment | @gemini-code-assist | Responds in comments when explicitly tagged, both in issue comments and review comments. |
| Help | /gemini help |
Displays a list of available commands. |
Customization
To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.
Limitations & Feedback
Gemini Code Assist is currently in preview and may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments to provide feedback.
You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.
Footnotes
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
There was a problem hiding this comment.
Code Review
The pull request introduces per-channel quantization for the KV cache. A potential issue was identified where tensor reshaping logic is applied unconditionally, which could break existing quantization strategies. The review suggests making this logic conditional to ensure the changes only affect the new per-channel quantization path.
| # reshape for per channel scenario | ||
| num_heads = key_states.shape[1] | ||
| head_dim = key_states.shape[-1] | ||
| # from [batch_size, num_heads, seq_len - residual_length, head_dim] | ||
| # to [batch_size, seq_len - residual_length, num_heads * head_dim] | ||
| key_states = key_states.transpose(1, 2).flatten(2) | ||
| value_states = value_states.transpose(1, 2).flatten(2) |
There was a problem hiding this comment.
The reshape logic is specific to per-channel quantization but is currently applied unconditionally. This can cause incorrect behavior for other quantization strategies. Consider adding a conditional check to apply this logic only when self.quantization_args.strategy == QuantizationStrategy.CHANNEL.
| # reshape for per channel scenario | |
| num_heads = key_states.shape[1] | |
| head_dim = key_states.shape[-1] | |
| # from [batch_size, num_heads, seq_len - residual_length, head_dim] | |
| # to [batch_size, seq_len - residual_length, num_heads * head_dim] | |
| key_states = key_states.transpose(1, 2).flatten(2) | |
| value_states = value_states.transpose(1, 2).flatten(2) | |
| # reshape for per channel scenario | |
| num_heads = key_states.shape[1] | |
| head_dim = key_states.shape[-1] | |
| if self.quantization_args.strategy == QuantizationStrategy.CHANNEL: | |
| # from [batch_size, num_heads, seq_len - residual_length, head_dim] | |
| # to [batch_size, seq_len - residual_length, num_heads * head_dim] | |
| key_states = key_states.transpose(1, 2).flatten(2) | |
| value_states = value_states.transpose(1, 2).flatten(2) |
| # reshape for per channel scenario | ||
| # from [batch_size, seq_len - residual_length, num_heads * head_dim] | ||
| # to [batch_size, num_heads, seq_len - residual_length, head_dim] | ||
| qdq_key_states = qdq_key_states.view( | ||
| qdq_key_states.shape[0], qdq_key_states.shape[1], num_heads, head_dim).transpose(1, 2) | ||
| qdq_value_states = qdq_value_states.view( | ||
| qdq_value_states.shape[0], qdq_value_states.shape[1], num_heads, head_dim).transpose(1, 2) |
There was a problem hiding this comment.
This reverse reshape operation is specific to the per-channel scenario and should be executed conditionally. Applying this to tensors from other quantization strategies will result in shape mismatches. Wrap this in a condition that checks for the per-channel strategy: if self.quantization_args.strategy == QuantizationStrategy.CHANNEL:.
| # reshape for per channel scenario | |
| # from [batch_size, seq_len - residual_length, num_heads * head_dim] | |
| # to [batch_size, num_heads, seq_len - residual_length, head_dim] | |
| qdq_key_states = qdq_key_states.view( | |
| qdq_key_states.shape[0], qdq_key_states.shape[1], num_heads, head_dim).transpose(1, 2) | |
| qdq_value_states = qdq_value_states.view( | |
| qdq_value_states.shape[0], qdq_value_states.shape[1], num_heads, head_dim).transpose(1, 2) | |
| if self.quantization_args.strategy == QuantizationStrategy.CHANNEL: | |
| # reshape for per channel scenario | |
| # from [batch_size, seq_len - residual_length, num_heads * head_dim] | |
| # to [batch_size, num_heads, seq_len - residual_length, head_dim] | |
| qdq_key_states = qdq_key_states.view( | |
| qdq_key_states.shape[0], qdq_key_states.shape[1], num_heads, head_dim).transpose(1, 2) | |
| qdq_value_states = qdq_value_states.view( | |
| qdq_value_states.shape[0], qdq_value_states.shape[1], num_heads, head_dim).transpose(1, 2) |
|
@Eviannn thank you for your PR! |
thx a lot! I wonder why test_cache.py uses 3 dims of key/value states while the annotate in QuantizedKVParameterCache shows that the shape is [batch_size, num_heads, seq_len - residual_length, head_dim] |
CI failed when running test_kv_cache.py, but in my local envs the script can pass...... |
SUMMARY:
kv cache quant int8 per channel is supported using this pr.
Besieds, compressed-tensors needs to be updated as well: neuralmagic/compressed-tensors#398
TEST PLAN:
specify type=int8 and strategy=channel in recipe