Remove permutation of KV cache #398
Open
+293
−271
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
This PR introduces a memory layout attribute for the MemoryType that allows the user to specify a physical shape that differs from the logical shape. This is useful in scenarios like kv-caches where certain dimensions physically are quite large but map to fixed logical dimensions. Signed-off-by: Harsh Menon <harsh@nod-labs.com>
harsh-nod
force-pushed
the
no_transposes
branch
from
d84210c to
f516332
Compare
harsh-nod
requested review from
raikonenfnu,
Hardcode84,
martin-luecke and
GMNGeoffrey
GMNGeoffrey
reviewed
Jan 21, 2025
lit_tests/kernel/wave/expansion.py
Outdated
| @@ -309,21 +309,21 @@ def test_gemm(): | |||
| # CHECK-SAME: (%b, 4, None, (), None) | |||
|
|
|||
| # CHECK-NEXT: %mma_M:0_N:0_K:0 | |||
| # CHECK-SAME: (%read_M:0_N:0_K:0, %read_M:0_N:0_K:0, %acc_M:0_N:0_K:0, None) | |||
| # CHECK-SAME: (%read_M:0_N:0_K:0, %read_M:0_N:0_K:0, %acc_M:0_N:0_K:0, None, None) | |||
There was a problem hiding this comment.
What if set_node_indices (or an earlier pass) actually wrote the MMA dimensional mapping to the node? Disadvantage: lit tests would be more verbose. Advantage: intermediate trace printing would make it much clearer what was going on. I was actually thinking that having all node attributes (so including vector shapes and indicies) attached to the nodes would be helpful.
There was a problem hiding this comment.
Doesn't the custom printing print the indices? (I don't think it print the vector shapes but we can add that as well)
Signed-off-by: Harsh Menon <harsh@nod-labs.com>
|
Turns out I didnt have to introduce a new dimensional mapping construct, this is already handled by mapping. |
harsh-nod
force-pushed
the
no_transposes
branch
from
f516332 to
d2fa3b3
Compare
This PR removes the permutation of the key and value cache prior to kernel invocation. Signed-off-by: Harsh Menon <harsh@nod-labs.com>
harsh-nod
force-pushed
the
no_transposes
branch
from
d2fa3b3 to
f5d1c07
Compare
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 PR removes the permutation of the key and
value cache prior to kernel invocation. In order to accomplish this, we add an additional attribute to the MMA operator that specifies which dimensions map to the M, N, K dimensions of the matrix multiplication.