This is an automated email from the ASF dual-hosted git repository.
tlopex pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/tvm.git
The following commit(s) were added to refs/heads/main by this push:
new 2516b1fd0d [REFACTOR][RUNTIME] Phase out IntTuple alias; use
ffi::Shape directly (#19471)
2516b1fd0d is described below
commit 2516b1fd0d96f61e4f47b2cd9cf163786b54facc
Author: Tianqi Chen <[email protected]>
AuthorDate: Wed Apr 29 14:46:57 2026 -0400
[REFACTOR][RUNTIME] Phase out IntTuple alias; use ffi::Shape directly
(#19471)
## Summary
`include/tvm/runtime/int_tuple.h` was a 2-line alias header
(`using IntTuple = ffi::Shape`). The alias adds indirection without
earning its keep — `ffi::Shape` is short enough that aliasing it gives
no ergonomic gain, and the layered name makes the reader follow one
extra
hop ("what's IntTuple? oh, Shape") at every call site.
- Deleted `include/tvm/runtime/int_tuple.h`
- Rewrote the 5 use sites (`disco/session.h`, `disco/bcast_session.h`,
`vm/kv_state.h`, `vm/attn_utils.h`, `vm/attn_backend.h`) to use
`ffi::Shape` / `ffi::ShapeObj` directly
- The `as_int_tuple()` member function in `attn_utils.h` keeps its name
(semantic identifier; renaming is out of scope)
---
include/tvm/runtime/disco/session.h | 4 ++--
include/tvm/runtime/int_tuple.h | 38 -------------------------------------
src/runtime/disco/bcast_session.h | 2 +-
src/runtime/vm/attn_backend.h | 1 -
src/runtime/vm/attn_utils.h | 7 ++++---
src/runtime/vm/kv_state.h | 15 ++++++++-------
6 files changed, 15 insertions(+), 52 deletions(-)
diff --git a/include/tvm/runtime/disco/session.h
b/include/tvm/runtime/disco/session.h
index daae2954af..c81a379454 100644
--- a/include/tvm/runtime/disco/session.h
+++ b/include/tvm/runtime/disco/session.h
@@ -73,8 +73,8 @@
#ifndef TVM_RUNTIME_DISCO_SESSION_H_
#define TVM_RUNTIME_DISCO_SESSION_H_
+#include <tvm/ffi/container/shape.h>
#include <tvm/ffi/function.h>
-#include <tvm/runtime/int_tuple.h>
#include <tvm/runtime/object.h>
#include <tvm/runtime/tensor.h>
@@ -240,7 +240,7 @@ class SessionObj : public Object {
* \param ccl The name of the communication backend, e.g., nccl, rccl, mpi.
* \param device_ids The device ids of the workers.
*/
- TVM_RUNTIME_DLL virtual void InitCCL(ffi::String ccl, IntTuple device_ids) =
0;
+ TVM_RUNTIME_DLL virtual void InitCCL(ffi::String ccl, ffi::Shape device_ids)
= 0;
/*!
* \brief Get the value of a register from a remote worker.
* \param reg_id The id of the register to be fetched.
diff --git a/include/tvm/runtime/int_tuple.h b/include/tvm/runtime/int_tuple.h
deleted file mode 100644
index 032e8b9a32..0000000000
--- a/include/tvm/runtime/int_tuple.h
+++ /dev/null
@@ -1,38 +0,0 @@
-/*
- * Licensed to the Apache Software Foundation (ASF) under one
- * or more contributor license agreements. See the NOTICE file
- * distributed with this work for additional information
- * regarding copyright ownership. The ASF licenses this file
- * to you under the Apache License, Version 2.0 (the
- * "License"); you may not use this file except in compliance
- * with the License. You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing,
- * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
- * KIND, either express or implied. See the License for the
- * specific language governing permissions and limitations
- * under the License.
- */
-
-/*!
- * \file int_tuple.h
- * \brief Defines tuple of integers.
- */
-#ifndef TVM_RUNTIME_INT_TUPLE_H_
-#define TVM_RUNTIME_INT_TUPLE_H_
-
-#include <tvm/ffi/container/shape.h>
-
-namespace tvm {
-namespace runtime {
-
-// We simply redirects to ffi::Shape, and ffi::ShapeObj
-using IntTuple = ffi::Shape;
-using IntTupleObj = ffi::ShapeObj;
-
-} // namespace runtime
-} // namespace tvm
-#endif // TVM_RUNTIME_INT_TUPLE_H_
diff --git a/src/runtime/disco/bcast_session.h
b/src/runtime/disco/bcast_session.h
index 119ca36409..dc9760ee81 100644
--- a/src/runtime/disco/bcast_session.h
+++ b/src/runtime/disco/bcast_session.h
@@ -41,7 +41,7 @@ class BcastSessionObj : public SessionObj {
void CopyToWorker0(const Tensor& host_array, const DRef& remote_array)
override;
void SyncWorker(int worker_id) override;
void Shutdown() override;
- void InitCCL(ffi::String ccl, IntTuple device_ids) override;
+ void InitCCL(ffi::String ccl, ffi::Shape device_ids) override;
ffi::Any DebugGetFromRemote(int64_t reg_id, int worker_id) override = 0;
void DebugSetRegister(int64_t reg_id, ffi::AnyView value, int worker_id)
override = 0;
diff --git a/src/runtime/vm/attn_backend.h b/src/runtime/vm/attn_backend.h
index 7f87f4f36d..5db83ff499 100644
--- a/src/runtime/vm/attn_backend.h
+++ b/src/runtime/vm/attn_backend.h
@@ -28,7 +28,6 @@
#include <tvm/ffi/container/array.h>
#include <tvm/ffi/function.h>
#include <tvm/runtime/device_api.h>
-#include <tvm/runtime/int_tuple.h>
#include <tvm/runtime/logging.h>
#include <memory>
diff --git a/src/runtime/vm/attn_utils.h b/src/runtime/vm/attn_utils.h
index afb962e4fc..c883705c82 100644
--- a/src/runtime/vm/attn_utils.h
+++ b/src/runtime/vm/attn_utils.h
@@ -24,6 +24,7 @@
#ifndef TVM_RUNTIME_VM_ATTN_UTILS_H_
#define TVM_RUNTIME_VM_ATTN_UTILS_H_
+#include <tvm/ffi/container/shape.h>
#include <tvm/runtime/tensor.h>
#include <algorithm>
@@ -286,7 +287,7 @@ GetBlockIdsOnDepth(const std::vector<Sequence*>& sequences,
* input blocks.
*/
inline std::pair<std::vector<std::pair<int32_t, int32_t>>, bool>
GetChunkedBlockIds(
- const std::vector<int32_t>& block_ids, bool enable_coalesce, const
IntTuple& append_lengths,
+ const std::vector<int32_t>& block_ids, bool enable_coalesce, const
ffi::Shape& append_lengths,
const std::vector<Block>& global_block_pool, bool is_decode_request) {
std::vector<std::pair<int32_t, int32_t>> uncoalesced_block_ids;
std::vector<std::pair<int32_t, int32_t>> coalesced_block_ids;
@@ -389,13 +390,13 @@ class HostMemoryVector {
/*! \brief Return the vector as an Tensor. */
Tensor as_tensor() { return data_.CreateView({current_size_}, data_->dtype);
}
- IntTuple as_int_tuple() const {
+ ffi::Shape as_int_tuple() const {
std::vector<int64_t> values;
values.reserve(current_size_);
for (int i = 0; i < current_size_; ++i) {
values.push_back(static_cast<int32_t*>(data_->data)[i]);
}
- return IntTuple(values);
+ return ffi::Shape(values);
}
private:
diff --git a/src/runtime/vm/kv_state.h b/src/runtime/vm/kv_state.h
index 33c669f18a..70a1fb8893 100644
--- a/src/runtime/vm/kv_state.h
+++ b/src/runtime/vm/kv_state.h
@@ -19,9 +19,10 @@
#ifndef TVM_RUNTIME_VM_KV_STATE_H_
#define TVM_RUNTIME_VM_KV_STATE_H_
#include <tvm/ffi/container/array.h>
+#include <tvm/ffi/container/shape.h>
#include <tvm/ffi/function.h>
+#include <tvm/ffi/optional.h>
#include <tvm/runtime/device_api.h>
-#include <tvm/runtime/int_tuple.h>
#include <tvm/runtime/logging.h>
#include <tvm/runtime/object.h>
#include <tvm/runtime/tensor.h>
@@ -95,8 +96,8 @@ class KVStateObj : public Object {
* is a chain.
*/
virtual void BeginForward(
- const IntTuple& seq_ids, const IntTuple& append_lengths,
- const ffi::Optional<IntTuple>& token_tree_parent_ptr = std::nullopt) = 0;
+ const ffi::Shape& seq_ids, const ffi::Shape& append_lengths,
+ const ffi::Optional<ffi::Shape>& token_tree_parent_ptr = std::nullopt) =
0;
/*!
* \brief Mark the start of the forward function.
@@ -155,15 +156,15 @@ class AttentionKVCacheObj : public KVStateObj {
* \param seq_ids The ids of the sequences to commit.
* \param leaf_indices The leaf token tree node index of each sequence.
*/
- virtual void CommitAcceptedTokenTreeNodes(const IntTuple& seq_ids,
- const IntTuple& leaf_indices) = 0;
+ virtual void CommitAcceptedTokenTreeNodes(const ffi::Shape& seq_ids,
+ const ffi::Shape& leaf_indices) =
0;
/*! \brief Prepare for the disaggregation KV data receive for the specified
sequence and length.*/
- virtual IntTuple DisaggPrepareRecv(int64_t seq_id, int length) = 0;
+ virtual ffi::Shape DisaggPrepareRecv(int64_t seq_id, int length) = 0;
/*! \brief Mark which tokens' KV cache needs to be sent to other devices */
virtual void DisaggMarkSend(int64_t seq_id, int64_t begin,
- const IntTuple& compressed_remote_position_map,
+ const ffi::Shape& compressed_remote_position_map,
int32_t recver_pe_offset) = 0;
/************** Attention **************/
