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 b4b97f8278 [Web] Allow custom bc files in emcc making (#16825)
add ffa9cfd0dd [BugFix][Ansor] Fixing Ansor Gradient Bug (#16739)
add 384b7f7a74 [KVCache] Introducing auxiliary data manager (#16824)
add b5fda2d93a [TIR] Ramp and Broadcast lanes fixed to int32 dtype (#16795)
add 3f615dcc3e [Bugfix][Relax] BlockBuilder may not assume unique input
functions (#16805)
add 00395ae43d [Relax][Bugfix] Provide the full Expr to pattern-match
rewriter (#16828)
add fc78b22fbc [Relax][VM] Refactor CUDA graph builtins as VM extension
(#16823)
add f83a32906f [Relax] Share storage allocs among functions after cuda
graph rewriting (#16830)
No new revisions were added by this update.
Summary of changes:
3rdparty/flashinfer | 2 +-
CMakeLists.txt | 2 +
include/tvm/runtime/relax_vm/vm.h | 44 ++
python/tvm/auto_scheduler/task_scheduler.py | 6 +
src/relax/ir/block_builder.cc | 38 +-
src/relax/ir/dataflow_matcher.cc | 13 +-
src/relax/transform/rewrite_cuda_graph.cc | 386 +++++++++++----
src/runtime/relax_vm/cuda/cuda_graph_builtin.cc | 60 ++-
src/runtime/relax_vm/paged_kv_cache.cc | 539 ++++++++++++++++-----
src/tir/ir/expr.cc | 8 +-
tests/python/arith/test_arith_rewrite_simplify.py | 15 +
.../test_relax_2d_buffer_allocation.py | 2 +
tests/python/relax/test_blockbuilder_core.py | 67 ++-
tests/python/relax/test_dataflow_pattern.py | 33 ++
..._builtin_paged_attention_kv_cache_flashinfer.py | 19 +-
...runtime_builtin_paged_attention_kv_cache_tir.py | 68 ++-
.../relax/test_transform_rewrite_cuda_graph.py | 241 ++++++++-
tests/python/tir-base/test_tir_nodes.py | 10 +
18 files changed, 1257 insertions(+), 296 deletions(-)