[cuda backend] int4/8 matvec: vectorized activation load #20144
Open
Gasoonjia wants to merge 2 commits into
Open
[cuda backend] int4/8 matvec: vectorized activation load #20144Gasoonjia wants to merge 2 commits into
Gasoonjia wants to merge 2 commits into
Conversation
…ock) The decode-only int4_plain_mm matvec was bound by activation load-instruction throughput, not DRAM bandwidth (already ~64% peak) or latency. Each inner iteration issued ~15 loads per 16-byte weight chunk: 8 scalar int32 activation loads + the same per-block scale d reloaded 4x. Align Q8Block to 16 bytes (sizeof 36->48) so each block's qs_even/qs_odd 16B halves are 16B-aligned, then load a whole activation block with two vectorized uint4 loads + one d load (~4x fewer activation loads). dp4a math and accumulation order are bit-identical; the int8 activation values and scale are unchanged. gemma4_31b decode (long-ctx harness, stacked on optimize_1): decode 43.98 -> 46.79 tok/s (+6.4%) prefill 1193 -> 1186 (noise; int4_plain_mm is decode-only) nsys: int4 matvec avg 38.4 -> 34.75 us (-9.5%); quant kernel unchanged. Unit tests test_aoti_torch_cuda_int4_plain_mm: 6/6 pass (M=1/8, gs=16/32/128).
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20144
Note: Links to docs will display an error until the docs builds have been completed. ⏳ 5 Pending, 2 Unrelated FailuresAs of commit 457a316 with merge base a79f3e4 ( FLAKY - The following jobs failed but were likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Gasoonjia
changed the title
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
The decode-only int4_plain_mm matvec was bound by activation load-instruction throughput, not DRAM bandwidth (already ~64% peak) or latency. Each inner iteration issued ~15 loads per 16-byte weight chunk: 8 scalar int32 activation loads + the same per-block scale d reloaded 4x. Same as int8_plain_mm
Align Q8Block to 16 bytes (sizeof 36->48) so each block's qs_even/qs_odd 16B halves are 16B-aligned, then load a whole activation block with two vectorized uint4 loads + one d load (~4x fewer activation loads). dp4a math and accumulation order are bit-identical; the int8 activation values and scale are unchanged.
gemma4_31b decode (long-ctx harness, stacked on optimize_1):
decode 43.98 -> 46.557 tok/s (+6.4%), +12.7% compare with llama.cpp (41.5 token/s)
profile result: int4 matvec avg 38.4 -> 34.75 us (-9.5%); quant kernel unchanged.