Add marlin int4 kernel #315
Closed
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
dacorvo
force-pushed
the
add_marlin_int4_kernel
branch
4 times, most recently
from
ffd984a to
5e5adbe
Compare
dacorvo
force-pushed
the
add_marlin_int4_kernel
branch
from
5e5adbe to
19ee33e
Compare
Original fix in vLLM project: The reason for the crash was the inline PTX assembly that introduced the async_copy with streaming behavior. The solution is to use the more standard PTX for async_copy (without the fractional L2 policy for "evict_first"). There is no performance difference between standard async_copy PTX and the previous one.
This is to guarantee Marlin kernels output is similar to the output obtained using dequantized weights.
dacorvo
force-pushed
the
add_marlin_int4_kernel
branch
from
19ee33e to
5564b8f
Compare
|
Closing this as the modified kernel is definitely flawed: as soon as it processes more than 32 inputs (i.e. two blocks of 16), there are errors in the outputs starting from the 128th output feature. There is likely some kind of flaw in the weight/scales/zero-point readback as soon as parallelization increases. |
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Outdated suggestions cannot be applied.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.
What does this PR do?
This adds a modified Marlin
fp16/int4kernel to the library and creates two newQTensorsubclasses to use it:MarlinInt4PackedTensor,MarlinInt4WeightQBitsTensor.The AWQ kernel is still used by default because from the first tests it seems the modified Marlin kernel either has some accuracy issues or is not properly integrated (perplexity increases).
Note: during the integration, I tried to register the Marlin fp16int4 gemm as a
torch.tensor.library.custom_op, but it added an extra latency (up to 50 %), so I used the same legacy declaration (withdefine/impl).