This is an automated email from the ASF dual-hosted git repository.
github-bot pushed a change to branch nightly
in repository https://gitbox.apache.org/repos/asf/tvm.git
from d91fe450c8 [Transform][Bugfix] Handle non-composite lambda functions
in FuseOps (#16598)
add 3ef478b486 [Relax][Runtime] RNNState for Space State Models (#16568)
add bd79374a01 [Bugfix][TVMScript] Handle R.match_cast as last binding in
if/else (#16562)
add ff0b99c5ce [Dlight] Scheduling Low batch GEMM using GEMV-like rule
(#16579)
No new revisions were added by this update.
Summary of changes:
python/tvm/dlight/gpu/__init__.py | 1 +
.../tvm/dlight/gpu/{gemv.py => low_batch_gemv.py} | 229 ++++++----
src/driver/driver_api.cc | 9 +-
src/runtime/relax_vm/kv_state.cc | 80 ++++
src/runtime/relax_vm/{kv_cache.h => kv_state.h} | 118 +++--
src/runtime/relax_vm/lm_support.cc | 11 +-
src/runtime/relax_vm/paged_kv_cache.cc | 41 +-
src/runtime/relax_vm/rnn_state.cc | 487 +++++++++++++++++++++
src/script/ir_builder/relax/frame.cc | 4 +-
src/script/ir_builder/relax/utils.h | 52 ++-
src/tir/transforms/hoist_expression.cc | 9 +-
tests/python/dlight/test_gpu_low_batch_gemv.py | 255 +++++++++++
.../python/relax/test_runtime_builtin_rnn_state.py | 262 +++++++++++
tests/python/relax/test_tvmscript_parser.py | 41 ++
14 files changed, 1434 insertions(+), 165 deletions(-)
copy python/tvm/dlight/gpu/{gemv.py => low_batch_gemv.py} (74%)
create mode 100644 src/runtime/relax_vm/kv_state.cc
rename src/runtime/relax_vm/{kv_cache.h => kv_state.h} (74%)
create mode 100644 src/runtime/relax_vm/rnn_state.cc
create mode 100644 tests/python/dlight/test_gpu_low_batch_gemv.py
create mode 100644 tests/python/relax/test_runtime_builtin_rnn_state.py
