[Kernel][ROCM] Upstream prefix prefill speed up for vLLM V1#13305

Merged

vllm-bot merged 73 commits intovllm-project:mainfrom

ROCm:upstream_prefix_prefill_speed_up

Apr 23, 2025

Commits on Feb 5, 2025

Commits on Feb 6, 2025

Commits on Feb 7, 2025

Commits on Feb 11, 2025

Commits on Feb 13, 2025

Commits on Feb 14, 2025

new prefix_prefill
root
committed

Commits on Feb 20, 2025

dwordx4 for k and v from cache
Aleksandr Malyshev
committed
merge with main
Aleksandr Malyshev
committed
follow up merge with main
Aleksandr Malyshev
committed

Commits on Feb 22, 2025

different stages for different loops
Aleksandr Malyshev
committed

Commits on Feb 24, 2025

merge with main
Aleksandr Malyshev
committed

Commits on Feb 25, 2025

unroll factors tunning
Aleksandr Malyshev
committed
linter
Aleksandr Malyshev
committed
default prefix_prefill for triton lower than 3.2, NV case
Aleksandr Malyshev
committed

Commits on Feb 28, 2025

Commits on Mar 4, 2025

original softmax restored to get back accuracy
root
committed
merge with main
Aleksandr Malyshev
committed

Commits on Mar 10, 2025

Merge branch 'upstream/main' into upstream_prefix_prefill_speed_up
Aleksandr Malyshev
committed
adaptation to ibm kernel
Aleksandr Malyshev
committed

Commits on Mar 11, 2025

softmax computation correction
Aleksandr Malyshev
committed

Commits on Mar 13, 2025

a comment for triton version
Aleksandr Malyshev
committed

Commits on Mar 19, 2025

Merge branch 'upstream_prefix_prefill_speed_up' of github.com:ROCm/vllm into upstream_prefix_prefill_speed_up
maleksan85
committed
Merge branch 'upstream/main' into upstream_prefix_prefill_speed_up
maleksan85
committed
kpack is not supported on NVidia triton
maleksan85
committed
kpack is not supported on NVidia triton
Aleksandr Malyshev
committed

Commits on Mar 27, 2025

reduced space of autotuning
Aleksandr Malyshev
committed

Commits on Apr 6, 2025

Merge branch 'upstream/main' into upstream_prefix_prefill_speed_up
Aleksandr Malyshev
committed

Commits on Apr 8, 2025

giving up on autotune and selecting one config
Aleksandr Malyshev
committed
Merge branch 'upstream/main' into upstream_prefix_prefill_speed_up
Aleksandr Malyshev
committed
fixing test with only to ROCM waves per eu and max_seq_len None
Aleksandr Malyshev
committed

Commits on Apr 9, 2025

renaming kernel
root
committed

Commits on Apr 10, 2025

clean up and fix for failed kernel tests
Aleksandr Malyshev
committed
clean up and fix for failed kernel tests
Aleksandr Malyshev
committed
clean up and fix for failed kernel tests
Aleksandr Malyshev
committed

Commits on Apr 11, 2025

got rid of autotuner and get stable runs right from the first iteration
maleksan85
committed

Commits on Apr 12, 2025

restoring paged attn as there is no autotuning anymore and that will no be error during start
maleksan85
committed

Commits on Apr 13, 2025

poking test rerun as one failed and seems not because of this change
maleksan85
committed

Commits on Apr 14, 2025

Merge branch 'main' of github.com:vllm-project/vllm into upstream_prefix_prefill_speed_up
maleksan85
committed
Merge branch 'upstream/main' into upstream_prefix_prefill_speed_up
maleksan85
committed
comment correction
maleksan85
committed

Commits on Apr 15, 2025

dot operation in triton doesn't support k to be 8 so increasing block size to most commonly used
maleksan85
committed
to kick CIs again Async Engine, Inputs, Utils, Worker Test seems flaky
maleksan85
committed
to kick CIs again
maleksan85
committed