nvfuser-cu121-torch25 0.2.25.dev20250201__cp310-cp310-manylinux_2_28_x86_64.whl

nvfuser/include/nvfuser/scheduler/multi_matmul.h

@@ -0,0 +1,74 @@
+// clang-format off
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2024-present NVIDIA CORPORATION & AFFILIATES.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+// clang-format on
+#pragma once
+
+#include <fusion.h>
+#include <scheduler/matmul_heuristic.h>
+#include <scheduler/mma_utils.h>
+#include <val_graph.h>
+#include <val_graph_visitor.h>
+#include <visibility.h>
+
+namespace nvfuser {
+
+// Base class for AmpereMultipleMatmulScheduler and
+// HopperMultipleMatmulScheduler
+class MultipleMatmulScheduler {
+ public:
+  MultipleMatmulScheduler(Fusion* fusion, const MatmulParams* params)
+      : fusion_(fusion),
+        params_(params),
+        id_model_(fusion, /*build_graphs=*/false) {}
+  virtual ~MultipleMatmulScheduler() = default;
+
+  virtual void run() = 0;
+
+ protected:
+  void findPatterns();
+
+  void translatePatterns();
+
+  // Get tensor roles and id roles
+  // When there are multiple matmul patterns, we can have conflicting roles.
+  // For now we throw an error if this is the case.
+  // TODO: This should be checked in canScheduleCompileTime
+  void findRoles();
+
+  void countDims();
+
+  //! Rebuilds IdModel, then updates all ValGroups in abstract tensors to refer
+  //! to the new IdModel. This is necessary whenever we perform an operation
+  //! that creates a new TensorView, such as caching or rFactor
+  void updateIdModel();
+
+ protected:
+  Fusion* fusion_;
+  const MatmulParams* params_;
+  IdModel id_model_;
+
+  // Broadcast graph of id_model_, which we modify at times using e.g.
+  // AbstractTensor.split or by mapping vals in cacheAfter and rFactor
+  ValGraph* graph_ = nullptr;
+  std::vector<mma_utils::MatmulPattern> patterns_;
+  mma_utils::DimRolesMap id_roles_;
+  mma_utils::TensorRolesMap tensor_roles_;
+  mma_utils::MatmulOperandInnerDims inner_dims_;
+
+  int64_t num_splitk_dims_ = 0;
+  int64_t num_device_dims_ = 0;
+  int64_t num_local_batch_dims_ = 0;
+  int64_t num_device_and_batch_dims_ = 0;
+
+  std::vector<TensorView*> as_, bs_, mma_results_;
+};
+
+NVF_API void scheduleMultipleMatmuls(
+    Fusion* fusion,
+    const MatmulParams* mparams);
+
+} // namespace nvfuser

nvfuser/include/nvfuser/scheduler/no_op.h

@@ -0,0 +1,48 @@
+// clang-format off
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2023-present NVIDIA CORPORATION & AFFILIATES.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+// clang-format on
+#pragma once
+
+#include <scheduler/heuristic.h>
+#include <scheduler/registry.h>
+
+namespace nvfuser {
+
+class Fusion;
+class SchedulerRuntimeInfo;
+class HeuristicDataCache;
+
+//! NoOp scheduler represents the case where scheduler will
+//!  not do any scheduling operations and forward the un-scheduled
+//!  fusion directly to code generation and kernel compilation.
+//!
+//! Typical use case of this scheduler is to handle edge cases
+//!  such as where all tensors are size-1 or size-0.
+
+class NoOpScheduler : public SchedulerEntry {
+ public:
+  //! Check if the no-op heuristics apply in given fusion
+  bool canScheduleCompileTime(Fusion* fusion) override;
+
+  bool canScheduleRunTime(
+      Fusion* fusion,
+      SchedulerRuntimeInfo& runtime_info,
+      HeuristicDataCache* data_cache = nullptr) override;
+
+  std::unique_ptr<HeuristicParams> computeHeuristics(
+      Fusion* fusion,
+      SchedulerRuntimeInfo& runtime_info,
+      HeuristicDataCache* data_cache) override;
+
+  void schedule(Fusion* fusion, const HeuristicParams* params) override;
+
+  constexpr static SchedulerType schedulerType() {
+    return SchedulerType::NoOp;
+  }
+};
+
+} // namespace nvfuser

nvfuser/include/nvfuser/scheduler/normalization_inner.h

@@ -0,0 +1,49 @@
+// clang-format off
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2023-present NVIDIA CORPORATION & AFFILIATES.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+// clang-format on
+#pragma once
+
+#include <ATen/core/ivalue.h>
+#include <exceptions.h>
+#include <fusion.h>
+#include <scheduler/reduction_heuristic.h>
+#include <scheduler/registry.h>
+#include <scheduler/utils.h>
+#include <visibility.h>
+
+// TODO: If caching inputs would require persistence we are sending it to the
+// persistent kerenl scheduler. This isn't necessary if the only persistent
+// buffers are inputs as we could re-read them from global memory. Need to
+// consider if this is worth implementing.
+
+namespace nvfuser {
+
+class SchedulerRuntimeInfo;
+class HeuristicDataCache;
+
+class InnerPersistentKernelScheduler : public SchedulerEntry {
+ public:
+  bool canScheduleCompileTime(Fusion* fusion) override;
+
+  bool canScheduleRunTime(
+      Fusion* fusion,
+      SchedulerRuntimeInfo& runtime_info,
+      HeuristicDataCache* data_cache = nullptr) override;
+
+  std::unique_ptr<HeuristicParams> computeHeuristics(
+      Fusion* fusion,
+      SchedulerRuntimeInfo& runtime_info,
+      HeuristicDataCache* data_cache) override;
+
+  void schedule(Fusion* fusion, const HeuristicParams* params) override;
+
+  constexpr static SchedulerType schedulerType() {
+    return SchedulerType::InnerPersistent;
+  }
+};
+
+} // namespace nvfuser

nvfuser/include/nvfuser/scheduler/normalization_inner_outer.h

@@ -0,0 +1,51 @@
+// clang-format off
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2023-present NVIDIA CORPORATION & AFFILIATES.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+// clang-format on
+#pragma once
+
+#include <ATen/core/ivalue.h>
+#include <exceptions.h>
+#include <fusion.h>
+#include <scheduler/reduction_heuristic.h>
+#include <scheduler/registry.h>
+#include <scheduler/utils.h>
+
+// TODO: If caching inputs would require persistence we are sending it to the
+// persistent kerenl scheduler. This isn't necessary if the only persistent
+// buffers are inputs as we could re-read them from global memory. Need to
+// consider if this is worth implementing.
+
+namespace nvfuser {
+
+class SchedulerRuntimeInfo;
+class HeuristicDataCache;
+
+class InnerOuterPersistentKernelScheduler : public SchedulerEntry {
+ public:
+  constexpr static int64_t threads_per_block_min = 128l;
+  constexpr static int64_t threads_per_block_max = 512l;
+
+  void schedule(Fusion* fusion, const HeuristicParams* params) override;
+
+  bool canScheduleCompileTime(Fusion* fusion) override;
+
+  bool canScheduleRunTime(
+      Fusion* fusion,
+      SchedulerRuntimeInfo& runtime_info,
+      HeuristicDataCache* data_cache = nullptr) override;
+
+  constexpr static SchedulerType schedulerType() {
+    return SchedulerType::InnerOuterPersistent;
+  }
+
+  std::unique_ptr<HeuristicParams> computeHeuristics(
+      Fusion* fusion,
+      SchedulerRuntimeInfo& runtime_info,
+      HeuristicDataCache* data_cache) override;
+};
+
+} // namespace nvfuser

nvfuser/include/nvfuser/scheduler/normalization_outer.h

@@ -0,0 +1,48 @@
+// clang-format off
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2023-present NVIDIA CORPORATION & AFFILIATES.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+// clang-format on
+#pragma once
+
+#include <ATen/core/ivalue.h>
+#include <exceptions.h>
+#include <fusion.h>
+#include <scheduler/reduction_heuristic.h>
+#include <scheduler/registry.h>
+#include <scheduler/utils.h>
+#include <visibility.h>
+
+// TODO: If caching inputs would require persistence we are sending it to the
+// persistent kerenl scheduler. This isn't necessary if the only persistent
+// buffers are inputs as we could re-read them from global memory. Need to
+// consider if this is worth implementing.
+
+namespace nvfuser {
+
+class SchedulerRuntimeInfo;
+class HeuristicDataCache;
+
+class OuterPersistentKernelScheduler : public SchedulerEntry {
+ public:
+  bool canScheduleCompileTime(Fusion* fusion) override;
+
+  bool canScheduleRunTime(
+      Fusion* fusion,
+      SchedulerRuntimeInfo& runtime_info,
+      HeuristicDataCache* data_cache = nullptr) override;
+
+  std::unique_ptr<HeuristicParams> computeHeuristics(
+      Fusion* fusion,
+      SchedulerRuntimeInfo& runtime_info,
+      HeuristicDataCache* data_cache) override;
+
+  void schedule(Fusion* fusion, const HeuristicParams* params) override;
+
+  constexpr static SchedulerType schedulerType() {
+    return SchedulerType::OuterPersistent;
+  }
+};
+} // namespace nvfuser

nvfuser/include/nvfuser/scheduler/normalization_utils.h

@@ -0,0 +1,379 @@
+// clang-format off
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2023-present NVIDIA CORPORATION & AFFILIATES.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+// clang-format on
+#pragma once
+
+#include <exceptions.h>
+#include <ir/all_nodes.h>
+#include <runtime/executor_params.h>
+#include <scheduler/reduction_utils.h>
+#include <scheduler/scheduler_types.h>
+#include <scheduler/utils.h>
+#include <cmath>
+#include <optional>
+#include <ostream>
+#include <vector>
+
+namespace nvfuser {
+class SchedulerRuntimeInfo;
+class HeuristicDataCache;
+
+namespace normalization_scheduler_utils {
+
+//! Utility class to iterate candidates of launch configurations in a
+//! preferred order. The iteration order is defined as:
+//!
+//!   for bdimx in all valid bdimx in an decreasing order
+//!     for gdimy in valid gdimy values in an increasing order
+//!
+//! Each of bdimx and gdimy determines bdimy and gdimx, respecitively,
+//! such that the number of threads per block is always 256 and the
+//! number of blocks is always equal to the number of SMs.
+class PreferredLaunchConfig {
+ public:
+  //! Minimum blockDim.x.
+  static constexpr int kMinBdimx = 8;
+  //! Maximum blockDim.x.
+  static constexpr int kMaxBdimx = 16;
+
+  PreferredLaunchConfig();
+
+  int bdimx() const {
+    return bdimx_;
+  }
+
+  int bdimy() const {
+    return bdimy_;
+  }
+
+  int gdimx() const {
+    return gdimxAt(grid_dims_pos_);
+  }
+
+  int gdimy() const {
+    return gdimyAt(grid_dims_pos_);
+  }
+
+  //! Peek the next gdimx. -1 is returned if no further gdimx is available.
+  int peekNextGdimx() const;
+
+  //! Peek the next gdimy. -1 is returned if no further gdimy is available.
+  int peekNextGdimy() const;
+
+  //! Move to the next launch configuration. Will be marked as invalid
+  //! if no valid configuration exists. Return true if successfully moved.
+  bool moveToNextConfig();
+
+  //! Try setting blockDim to the next valid config if
+  //! available. Return false if no valid config exists. gridDim is
+  //! reset.
+  bool moveToNextBdim();
+
+  //! Query if the next configuration will cause blockDim.x to become
+  //! smaller.
+  bool isNextSmallerBdimx() const;
+
+  //! Query if blockDim.x can be further lowered
+  bool canLowerBdimx() const;
+
+  //! Query if no valid configuration is found
+  bool isInvalid() const {
+    return !valid_;
+  }
+
+ private:
+  //! Populate the list of valid gridDim configurations
+  void initValidGdims();
+
+  int gdimxAt(int pos) const {
+    return valid_grid_dims_.at(pos).first;
+  }
+
+  int gdimyAt(int pos) const {
+    return valid_grid_dims_.at(pos).second;
+  }
+
+  //! Set blockDim.x and in turn blockDim.y. Return true if the
+  //! specified blockDim.x is successfully set. If dry_run is true,
+  //! just check if the given config is valid but do not modify the
+  //! current config.
+  bool setBdimx(int bdimx, bool dry_run = false);
+
+  void resetGdim() {
+    grid_dims_pos_ = 0;
+  }
+
+  void resetBdim() {
+    // Start with the maximum bdimx and lower it until satisfactory
+    // config is found
+    setBdimx(kMaxBdimx);
+  }
+
+  //! Try setting gridDim to the next valid config if
+  //! available. Return false if no valid config exists
+  bool moveToNextGdim();
+
+  int getNextGdimsPos() const;
+
+  void invalidate() {
+    valid_ = false;
+  }
+
+  friend std::ostream& operator<<(std::ostream& os, PreferredLaunchConfig cfg) {
+    os << "{gdimx: " << cfg.gdimx() << ", gdimy: " << cfg.gdimy()
+       << ", bdimx: " << cfg.bdimx() << ", bdimy: " << cfg.bdimy() << "}";
+    return os;
+  }
+
+ private:
+  //! Remember if it is still a valid configuration
+  bool valid_ = false;
+
+  //! List of valid gridDims ordered by the dimension of
+  //! gridDim.x. Larger gridDim.x is preferred as it would promote
+  //! larger independent parallelism
+  std::vector<std::pair<int, int>> valid_grid_dims_;
+  //! The offset of the Current gridDim in valid_grid_dims_
+  int grid_dims_pos_ = 0;
+
+  //! Current blockDim.x
+  int bdimx_ = 0;
+  //! Current blockDim.y
+  int bdimy_ = 0;
+};
+
+//! Scheduling parameters for grid outer normalization
+struct GridOuterNormalizationParams {
+  LaunchParams launch_params;
+  int64_t persistent_buffer_factor = -1;
+  int64_t unswitch_factor = -1;
+};
+
+std::optional<GridOuterNormalizationParams> getGridOuterNormalizationParams(
+    int64_t total_reduction_numel,
+    int64_t total_iteration_numel,
+    int64_t vectorize_factor,
+    int64_t persistent_buffer_size);
+
+//! check iter type of each domain in inner and outer reduction tvs
+//! inner reduction must be [I,I,...R,R]
+//! outer reduction must be [R,R,...I,I]
+bool checkIfReductionsAreInnerOuter(
+    const std::vector<TensorView*>& inner_reduction_tvs,
+    const std::vector<TensorView*>& outer_reduction_tvs);
+
+//! check if the inner reduction has shared input with outer reduction
+bool hasSharedInput(
+    const std::vector<TensorView*>& inner_reduction_tvs,
+    const std::vector<TensorView*>& outer_reduction_tvs);
+
+//! The first part of outer reduction is computed with inner reduction and the
+//! second part is scheduled separately. So, (1) the outer reduction tvs can
+//! only be connected with inner reduction tvs through their producers. (2)
+//! Outer reduction tvs are also scheduled separately and they can only be
+//! connected through their producers.
+bool isConnectedOnlyThroughReductionProducer(
+    const std::vector<TensorView*>& inner_reduction_tvs,
+    const std::vector<TensorView*>& outer_reduction_tvs);
+
+// Returns true if every iteration domain in inner reduction tv is a reduction
+// domain in outer reduction tv.
+bool isReductionIterationAxisMatched(
+    const std::vector<TensorView*>& inner_reduction_tvs,
+    const std::vector<TensorView*>& outer_reduction_tvs);
+
+//! in combined_inner_outer_reduction, the partial results of outer reductions
+//! must be persistent, calculate the size of these buffers when estimate
+//! register usage
+int64_t partialReductionBufferSize(
+    const std::vector<TensorView*>& outer_reduction_tvs,
+    SchedulerRuntimeInfo& runtime_info);
+
+// Return a scheduleHeuristic based on reduction types.
+using ReductionType = reduction_scheduler_utils::ReductionType;
+SchedulerType getPersistentHeuristicFor(ReductionType reduction_type);
+
+struct PersistentKernelProperties {
+  int64_t inner_most_dimension_numel;
+  int64_t total_reduction_numel;
+  int64_t total_iteration_numel;
+  int64_t max_persistent_buffer_size;
+  int64_t n_tensor_inputs;
+  int64_t max_dtype_size;
+  int64_t vectorize_factor;
+  bool project_persistent_buffers;
+  PrimDataType index_type;
+  bool has_exp_op;
+  bool has_rng_op;
+  bool disable_project_to_avoid_recompute;
+  std::vector<TensorView*> persistent_buffers;
+  std::string toString() const {
+    std::stringstream ss;
+    ss << "===== Persistent Kernel Properties ========\n"
+       << "inner_most_dimension_numel: " << inner_most_dimension_numel << "\n"
+       << "total_reduction_numel: " << total_reduction_numel << "\n"
+       << "total_iteration_numel: " << total_iteration_numel << "\n"
+       << "max_persistent_buffer_size: " << max_persistent_buffer_size << "\n"
+       << "n_tensor_inputs: " << n_tensor_inputs << "\n"
+       << "max_input_dtype_size: " << max_dtype_size << "\n"
+       << "max allowed vectorize_factor: " << vectorize_factor << "\n"
+       << "disable_project_to_avoid_recompute: "
+       << disable_project_to_avoid_recompute << "\n"
+       << "project_persistent_buffers: " << project_persistent_buffers << "\n";
+    return ss.str();
+  }
+};
+PersistentKernelProperties getPersistentKernelProperties(
+    Fusion* fusion,
+    SchedulerRuntimeInfo& runtime_info,
+    HeuristicDataCache* data_cache,
+    SchedulerType heuristic);
+
+// Verify the presence of a reduction TensorView connected to a Fusion input
+void checkReductionTvForScheduling(Fusion* fusion, TensorView* ref_red_tv);
+
+// Check the operations and input tensors of the fusion. This
+// verification is a common step shared by all persistent kernel implementations
+// during compile-time checks.
+bool checkOpsAndInputs(Fusion* fusion, SchedulerType scheduler_type);
+
+// Returns true if the reduction pattern is consistent. For the
+// InnerPersistentKernelScheduler and OuterPersistentKernelScheduler, a single
+// vector of TensorViews is provided, while for the
+// InnerOuterPersistentKernelScheduler, two vectors of TensorViews are provided.
+bool checkReductionPattern(
+    Fusion* fusion,
+    SchedulerType scheduler_type,
+    const std::vector<TensorView*>& reduction_tvs1,
+    const std::vector<TensorView*>& reduction_tvs2 = {});
+
+// The compile-time checks for both the InnerPersistentKernelScheduler and
+// OuterPersistentKernelScheduler are identical. These checks are constructed
+// using checkOpsAndInputs, checkReductionPattern, and checkViewBufferTopology.
+bool compileTimeCheck(Fusion* fusion, SchedulerType scheduler_type);
+
+// Common preparations before the actual schedule, used by all persistent
+// schedulers. Write to dummy_outputs, cached_inputs, reduction_tvs, and
+// cached_outputs.
+void beforeSchedule(
+    Fusion* fusion,
+    const ReductionParams* rparams,
+    std::vector<TensorView*>& dummy_outputs,
+    std::vector<TensorView*>& cached_inputs,
+    std::vector<TensorView*>& reduction_tvs,
+    std::vector<TensorView*>& smem_consumers,
+    std::vector<std::pair<TensorView*, TensorView*>>& cached_outputs);
+
+// schedule a reduction tv, used by all persistent schedulers.
+// will group reduction ops for OuterPersistentKernelScheduler with multiple
+// reduction tvs.
+TensorView* scheduleReductionGeneral(
+    Fusion* fusion,
+    const ReductionParams* rparams,
+    std::vector<TensorView*>& reduction_tvs,
+    SchedulerType scheduler_type);
+
+// Used by InnerPersistentKernelScheduler and  OuterPersistentKernelScheduler
+void schedulePersistentKernel(
+    Fusion* fusion,
+    const ReductionParams* rparams,
+    SchedulerType scheduler_type);
+
+// Get max register or shared memory size for persistent buffer
+int64_t getMaxRegOrSharedMemorySizeForPersistentBuffer(
+    SchedulerRuntimeInfo& runtime_info,
+    const std::vector<TensorView*>& persistent_buffers,
+    const bool can_use_smem_persistent);
+
+enum class BufferProjectionStrategy {
+  // Recompute persistent buffers from inputs, only need to cache inputs in
+  // registers or shared memories, usually used when size of required cached
+  // inputs is smaller than the size of persistent buffers.
+  ProjectToInputs,
+  // Don't project to inputs, to avoid recompute from inputs. This saves
+  // computation cost but uses more registers or shared memories. Usually used
+  // when the required buffer size  is small and hardware has high bandwidth to
+  // flops ratio.
+  NoProjectToAvoidRecompute,
+  // Project to inputs is disabled due to other reasons, e.g. can't reduce
+  // buffer size, recompute requires very expensive rng ops, not supported due
+  // to view ops.
+  NoProjectOtherReasons
+};
+
+// Returns BufferProjectionStrategy based on buffer size, hardware, and fusion
+// ops.
+
+// This function is used by inner persistent and InnerOuter persistent
+// schedulers.
+// Using shared memory to store persistent buffers is not supported yet for
+// inner persistent scheduler with 3D reduction type.
+// TODO: Outer persistent scheduler should also use this function.
+// If the scheduler is innerOuter with outer broadcast, projection is allowed
+// even it leads to a larger buffer size becuase the scheduled kernel allows the
+// reuse of the outer broadcast Tv when iterating over the outer reduction
+// dimension and leads to higher performance ( TODO: needs re-evaluate, may not
+// true if the buffer size is increased a lot when projecting to inputs). See
+// https://github.com/NVIDIA/Fuser/issues/402
+
+// However, we experimentally found that certain relatively expensive operations
+// should not be projected even when that would require a larger buffer size.
+// Specifically,
+// - rng: should never be projected no matter how much larger the buffer would
+// consume
+// - exp in inner normalization: only allowed to get projected if the buffer is
+// smaller than a certain size Otherwise, as long as the projected inputs are
+// smaller than the original persistent buffers, this function returns true.
+BufferProjectionStrategy isProjectBufferToInputs(
+    Fusion* fusion,
+    SchedulerRuntimeInfo& runtime_info,
+    const scheduler_utils::PersistentBufferInfo& persistent_buffer_info,
+    const scheduler_utils::PersistentBufferSizeReturn&
+        persistent_buffer_size_info,
+    const SchedulerType sh,
+    const bool can_use_smem_persistent,
+    const bool check_projected_buffer_size = true);
+
+// Set memory type of persistent buffer marked in
+// rparams->smem_persistent_buffers as shared memory. Return a vector of the
+// consumers of the shared memory tensors, they are cached after the smem
+// tensors and will be vectorized by the scheduler if possible to avoid shared
+// memory bank conflicts.
+std::vector<TensorView*> movePersistentBufferToSmem(
+    Fusion* fusion,
+    const ReductionParams* rparams,
+    const std::vector<TensorView*>& cached_inputs);
+
+// Find the resolution points of a persistent buffer. See also
+// the comments of PersistentBufferResolution in utils.cpp. Unlike
+// PersistentBufferResolution, this analysis traverses a given fusion
+// both forward and backward, which is necessary in some cases. For
+// example:
+//
+// t0 = makeSymbolicTensor(2)
+// t1 = makeSymbolicTensor(2)
+// t2 = set(t0)
+// t3 = sum(t2, 1)
+// t4 = broadcast(t3, {false, true})
+// t5 = add(t1, t2)
+// t6 = add(t4, t1)
+// fusion.addOutput(t5)
+// fusion.addOutput(t6)
+//
+// The path from t2 to t3, t4 and t6 is a normalization path. While t1 itself
+// does not depend on t2, since it is used with t2, inlining of t2
+// also means t1 must be inlined, which in turn means t6 must be
+// inlined. However, t6 depends on the reduction, inlining of t2 is
+// not possible. For normalization fusions like this pattern,
+// PersistentBufferResolution is not able to detect the resolution
+// point. getResolutionPointsOf addresses the problem by traversing
+// both forward and backward directions. See
+// PersistentBufferTest.GetResolutionIssue1123 for a concrete example
+std::vector<TensorView*> getResolutionPointsOf(TensorView* persistent_buffer);
+
+} // namespace normalization_scheduler_utils
+} // namespace nvfuser