Convolutional Coded Quantization

[tsengalb99]

I recently came across your CCQ quantization method as described in https://arxiv.org/pdf/2507.07145, and I'm have some questions about the coding setup. My understanding is that the method does the following

1. First encode vectors of dimension 3 with a (L=2, N=3, S=2) and vectors of dimension 4 with a (3, 4, 2) setup to get a compressed bitstream and set of scales
2. Further compress the bitstreams by doing vector quantization to a "meta codebook"
3. Compress the scales

My questions are:

1. The baseline encoding scheme appears to just be QTIP where L = L, N = the vector dimension, V = 1, and S = K = the bitrate. The codebook is a direct lookup (bit to integer) codebook with a scale. The bitshift decoding section in your paper corresponds to the bitshift trellis in QTIP. Can you clarify how the base encoding scheme is fundamentally different from QTIP and not just a variant of QTIP?
2. If my understanding is correct, does step 2 cluster the 3D vectors together and the 4D vectors together, or the bitstream together as a 7D vector?
3. The end result of this seems to be a 4D + 3D vector quantizer or 7D VQ. If its 4D and 3D (two separate vector quantizers), why not just use unstructured 3D and 4D VQ? The codebook should be small enough to fit in cache in both cases.
4. Why not use entropy coding since in step 2 instead of higher order vector quantization?
5. Is the rate distortion tradeoff of doing step 2 better than just doing TCQ at the target bitrate?
6. What is the "base" rounding algorithm? The paper states that within each vector in step 1, you find the minimize MSE representation, which corresponds to rounding to the nearest. However, RTN is known to do very poorly regardless of how good your quantizer is. Do you use GPTQ, AWQ, or something else?

I would have sent an email to the authors except the paper doesn't list the author emails.

Thanks

[zzjjay]

Thank you for your attention. Our CCQ algorithm is precisely inspired by QTIP (which is an outstanding work), leading us to focus on the application of coding algorithms in quantization.

Regarding the several questions you raised, here are detailed explanations:
1.Differences from QTIP: The fundamental coding concepts are similar, as both originate from convolutional codes. However, unlike QTIP, which encodes long sequences, our approach focuses on constructing linear codebooks for short vectors. This enables us to employ vector search for encoding purposes, rather than relying on the Viterbi algorithm, thereby enhancing coding efficiency.
2.The aggregation of 3D and 4D vectors into a 7D vector that you mentioned is part of our hybrid coding configuration. This configuration is designed to achieve 2.28 bpw. Under quantization at 2 bpw, we employ code clustering to further compress the codebook space, rather than simply aggregating 3D and 4D vectors.
3.We refrain from using unstructured codebooks to eliminate the dependency on them during decoding, thereby transforming the decoding process entirely into pure computational operations and enhancing computational efficiency. Even though a smaller codebook can fit into the cache, it would introduce asynchronous operations that could adversely affect inference efficiency.
4.Similar to the previous issue, employing entropy coding introduces a reliance on the codebook, which runs counter to our original intention.
5.Directly employing convolutional code presents a trade-off between storage requirements and performance. Given our intention to utilize a relatively high number of state bits (i.e., L), it becomes challenging to maintain code values that precisely fit into storage formats such as int8 or int16 while achieving a high compression ratio.
6.To mitigate the impact of RTN under low-bit conditions, we employ a relatively high number of state bits as much as possible on the basis of group-wise quantization. We also explored integrating it with GPTQ; however, experimental results demonstrated that using GPTQ offered limited benefits for CCQ and even reduced the algorithm's efficiency.

I sincerely hope that my answers can clarify your doubts.