CUDA: Enable K-shift operation for -ctk q8_0 (limited) #9571
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
I cannot reproduce the issue that you reported with memory usage, in my tests the allocator correctly reuses the memory of the previous tensors automatically. The It would also be very desirable to implement this in the CPU backend, and add a test in For cases where the input and outputs of the copy are contiguous (as it is here), then this could also be implemented using the existing dequantize functions in the CUDA backend, which would allow it to work with any format and likely with better performance. |
Nekotekina
force-pushed
the
kshift
branch
2 times, most recently
from
ea88404
to
eec216c
Compare
|
Right, I simplified it. Thanks. But where are dequantize functions you are talking about? You mean |
|
I also forgot to ask, is there a reason to use f32 over f16? The performance of q8_0 K-Shift doesn't seem great in this PR, I wonder if using f16 can improve it... |
|
Yes, I mean the functions from |
slaren
approved these changes
Sep 22, 2024
llama: enable K-shift for quantized KV cache It will fail on unsupported backends or quant types.
Great, it works well in actual use. Is there a plan to support more quantitative specifications, such as q4_0? |
|
I'd like to improve it more, as @slaren mentioned, but I'm confused a bit by APIs. If I use convert function like this, what should be in |
|
Yes, that looks right. |
|
Hmm, |
dsx1986
pushed a commit
to dsx1986/llama.cpp
that referenced
this pull request
Oct 29, 2024
llama: enable K-shift for quantized KV cache It will fail on unsupported backends or quant types.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This is a reworked #5653
Some CUDA code was adapted from: 3d92acf
Original PR had explosive GPU memory requirement.
I'm not sure if it's a bug, or intended logic of the allocator.I worked around it by reusing the same tensor. It seems to work well for me.