Remove legacy bindings

[rdspring1]

This PR removes the nvfuser python module, corresponding pybind11 CPP bindings, and any references from csrc. Version is bumped to 0.2.36.

[rdspring1]

!test

[github-actions]

Review updated until commit 11be6d7

Description

• Removes legacy nvfuser python module and pybind11 bindings from build system

• Eliminates all source files and configuration for legacy python bindings

• Adds NVF_API to function declarations in type.h and multidevice/utils.h

• Updates version to 0.2.36 and modifies documentation to reflect removed serialization

Changes walkthrough

	Relevant files
Enhancement	5 files CMakeLists.txt Remove legacy nvfuser python library build configuration  +3/-141  utils.py Remove nvfuser._C extension handling, keep only direct version +1/-5      version.txt Bump version from 0.2.35 to 0.2.36                                              +1/-1      type.h Add NVF_API to UnaryOpType and TernaryOpType operator<< declarations +2/-2      utils.h Add NVF_API to getShardedLogicalAxis function declaration +2/-1
Documentation	1 files Serde.md Add note about serialization being disabled due to removed legacy bindings +3/-1
Miscellaneous	1 files fusion_segmenter.cpp Update test reference URL to new location in direct tests +1/-1
Additional files	34 files options.cpp +0/-2      options.h +0/-2      fusion_kernel_runtime.cpp +0/-12    fusion_record.cpp +0/-952  fusion_record.h +0/-124  README.md +0/-210  __init__.py +0/-649  __init__.pyi +0/-4      benchmark_utils.py +0/-160  __init__.py +0/-9      __init__.py +0/-13    normalization.py +0/-725  nvfuser_version.py +0/-69    pytorch_utils.py +0/-190  fusion_cache.cpp +0/-953  fusion_cache.h +0/-320  fusion_definition.cpp +0/-769  fusion_definition.h +0/-389  fusion_record.h +0/-3675 fusion_state.cpp +0/-297  fusion_state.h +0/-143  multidevice_bindings.cpp +0/-103  python_bindings.cpp +0/-4196 python_bindings.h +0/-27    python_bindings_extension.cpp +0/-18    schedule_bindings.cpp +0/-517  segmentation.cpp +0/-369  segmentation.h +0/-246  translation.cpp +0/-1484 translation.h +0/-20    translation_utils.cpp +0/-80    translation_utils.h +0/-300  test_import.py +0/-17    env_options.yaml +0/-12

PR Reviewer Guide

Here are some key observations to aid the review process:

🧪 PR contains tests

⚡ Recommended focus areas for review

Build System Changes The build_extension method now only handles nvfuser_direct._C_DIRECT extension, removing support for the legacy nvfuser._C extension. This is a significant change that affects the build system and users who might still depend on the old extension name. if ext.name == "nvfuser_direct._C_DIRECT": self.copy_library(ext, "libnvfuser_direct") self.copy_shared_library("libnvfuser_codegen.so") else: CMake Build Configuration The entire nvfuser python library build section has been removed from CMakeLists.txt. This includes the python API version generation, library compilation, and installation rules. Need to verify this doesn't break existing build workflows or documentation. # ------------------------------------------------ # build nvfuser direct python library # ------------------------------------------------ # nvfuser direct bindings API sources set(NVFUSER_PYTHON_DIRECT_SRCS) list(APPEND NVFUSER_PYTHON_DIRECT_SRCS ${NVFUSER_PYTHON_DIRECT_BINDINGS}/extension.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/bindings.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/enum.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/heuristic_params.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/ir.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/internal_ir.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/lru_cache.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/multidevice.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/ops.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/cutlass.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/runtime.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/schedule.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/id_model.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/profile.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/direct_utils.cpp ${NVFUSER_PYTHON_DIRECT_BINDINGS}/python_translate.cpp ${NVFUSER_PYTHON_COMMON}/distributed_tensor.cpp ${NVFUSER_PYTHON_COMMON}/python_utils.cpp ${NVFUSER_PYTHON_COMMON}/translation_names.cpp ) add_library(nvf_py_direct_internal OBJECT ${NVFUSER_PYTHON_DIRECT_SRCS}) # setup python API version add_custom_command( OUTPUT ${NVFUSER_PYTHON_DIR}/nvfuser_direct/version.py COMMAND "${Python_EXECUTABLE}" -c \"from pathlib import Path\; Path('${NVFUSER_PYTHON_DIR}/tools/gen_nvfuser_version.py') .touch() \" COMMAND "${Python_EXECUTABLE}" ${NVFUSER_PYTHON_DIR}/tools/gen_nvfuser_version.py nvfuser_direct DEPENDS ${NVFUSER_PYTHON_DIR}/tools/gen_nvfuser_version.py DEPENDS ${NVFUSER_PYTHON_DIR}/version.txt WORKING_DIRECTORY ${NVFUSER_PYTHON_DIR}/tools/ ) add_custom_target( gen_nvfuser_direct_version ALL DEPENDS ${NVFUSER_PYTHON_DIR}/nvfuser_direct/version.py ) add_dependencies(nvf_py_direct_internal gen_nvfuser_direct_version) # NOTE: For any future extension, change PYTHON_DIRECT_EXTENSION to another # name other than EXTENSION_NAME. target_compile_definitions(nvf_py_direct_internal PRIVATE "-DTORCH_CUDA_BUILD_MAIN_LIB" "-DC10_BUILD_MAIN_LIB=1" PYTHON_DIRECT_EXTENSION=_C_DIRECT ) add_library(nvfuser_direct MODULE $<TARGET_OBJECTS:nvf_py_direct_internal>) target_compile_definitions(nvfuser_direct PRIVATE "-DTORCH_CUDA_BUILD_MAIN_LIB" "-DC10_BUILD_MAIN_LIB=1" PYTHON_DIRECT_EXTENSION=_C_DIRECT ) if(NOT MSVC) target_compile_options(nvf_py_direct_internal PRIVATE -Wall -Wno-unused-function) target_compile_options(nvf_py_direct_internal PRIVATE -Werror) # Add function/data sections for dead code elimination target_compile_options(nvf_py_direct_internal PRIVATE "-ffunction-sections" "-fdata-sections" ) set(NVF_LIB_SUFFIX ".so") else() set(NVF_LIB_SUFFIX ".pyd") endif() set_target_properties(nvf_py_direct_internal PROPERTIES C_STANDARD ${NVFUSER_C_STANDARD} CUDA_STANDARD ${NVFUSER_CUDA_STANDARD} CXX_STANDARD ${NVFUSER_CPP_STANDARD} CXX_STANDARD_REQUIRED ON CXX_VISIBILITY_PRESET hidden INSTALL_RPATH "$ORIGIN/lib:$ORIGIN/../nvidia/cuda_runtime/lib:$ORIGIN/../nvidia/cuda_nvrtc/lib:$ORIGIN/../../nvidia/cuda_cupti/lib:$ORIGIN/../torch/lib" POSITION_INDEPENDENT_CODE Yes VISIBILITY_INLINES_HIDDEN Yes ) set_target_properties(nvfuser_direct PROPERTIES C_STANDARD ${NVFUSER_C_STANDARD} CUDA_STANDARD ${NVFUSER_CUDA_STANDARD} CXX_STANDARD ${NVFUSER_CPP_STANDARD} CXX_STANDARD_REQUIRED ON CXX_VISIBILITY_PRESET hidden INSTALL_RPATH "$ORIGIN/lib:$ORIGIN/../nvfuser_common/lib:$ORIGIN/../nvidia/cuda_runtime/lib:$ORIGIN/../nvidia/cuda_nvrtc/lib:$ORIGIN/../../nvidia/cuda_cupti/lib:$ORIGIN/../torch/lib" POSITION_INDEPENDENT_CODE Yes SUFFIX ${NVF_LIB_SUFFIX} VISIBILITY_INLINES_HIDDEN Yes ) target_include_directories(nvf_py_direct_internal PUBLIC ${NVFUSER_PYTHON_DIRECT_BINDINGS}) target_include_directories(nvf_py_direct_internal PUBLIC ${NVFUSER_PYTHON_COMMON}) target_link_libraries(nvf_py_direct_internal PRIVATE nvfuser_codegen "${TORCH_INSTALL_PREFIX}/lib/libtorch_python.so" pybind11::pybind11 pybind11::headers CUDA::cupti ) if (NVFUSER_USE_CUTLASS) target_link_libraries(nvf_py_direct_internal PRIVATE nvf_cutlass) endif() target_link_libraries(nvfuser_direct PRIVATE nvf_py_direct_internal Python::Module ) # Add dead code elimination flags to reduce file size if(NOT MSVC) target_link_options(nvfuser_direct PRIVATE "-Wl,--gc-sections" "-Wl,--as-needed" $<$<CONFIG:Release>:-s> ) endif() set_target_properties(nvfuser_direct PROPERTIES INSTALL_RPATH "$ORIGIN:$ORIGIN/../build:$ORIGIN/../nvfuser_common/lib" ) install(TARGETS nvfuser_direct DESTINATION lib) endif() set(JIT_TEST_SRCS) list(APPEND JIT_TEST_SRCS ${NVFUSER_ROOT}/tests/cpp/kernel_db/test_nvfuser_kernel_db_open.cpp ${NVFUSER_ROOT}/tests/cpp/kernel_db/test_nvfuser_kernel_db_query.cpp ${NVFUSER_ROOT}/tests/cpp/kernel_db/test_nvfuser_kernel_db_write.cpp ${NVFUSER_ROOT}/tests/cpp/test_abstract_tensor.cpp ${NVFUSER_ROOT}/tests/cpp/test_alias.cpp ${NVFUSER_ROOT}/tests/cpp/test_alias_analysis.cpp ${NVFUSER_ROOT}/tests/cpp/test_allocation_domain.cpp ${NVFUSER_ROOT}/tests/cpp/test_allocation_order_inference.cpp ${NVFUSER_ROOT}/tests/cpp/test_bfs.cpp ${NVFUSER_ROOT}/tests/cpp/test_ca_root_domain_map.cpp ${NVFUSER_ROOT}/tests/cpp/test_circular_buffering.cpp ${NVFUSER_ROOT}/tests/cpp/test_circular_buffering_ping_pong.cpp ${NVFUSER_ROOT}/tests/cpp/test_combined_inner_outer_reduction.cpp ${NVFUSER_ROOT}/tests/cpp/test_compute_at_map.cpp ${NVFUSER_ROOT}/tests/cpp/test_compute_with.cpp ${NVFUSER_ROOT}/tests/cpp/test_contiguity_id_model.cpp ${NVFUSER_ROOT}/tests/cpp/test_driver_api.cpp ${NVFUSER_ROOT}/tests/cpp/test_dynamic_transform.cpp ${NVFUSER_ROOT}/tests/cpp/test_embedding_node.cpp ${NVFUSER_ROOT}/tests/cpp/test_evaluator.cpp ${NVFUSER_ROOT}/tests/cpp/test_exceptions.cpp ${NVFUSER_ROOT}/tests/cpp/test_expr_simplifier.cpp ${NVFUSER_ROOT}/tests/cpp/test_expr_sort.cpp ${NVFUSER_ROOT}/tests/cpp/test_fusion_hash.cpp ${NVFUSER_ROOT}/tests/cpp/test_gather.cpp Runtime Dependencies Removed includes for python_frontend headers and debug code for PythonDefinitionSegments. This changes the runtime dependencies and removes debugging capabilities. Need to ensure this doesn't affect runtime functionality or debugging capabilities for users. #include <c10/cuda/CUDAGuard.h> #include "fusion.h" #include "fusion_profiler.h" #include "fusion_segmenter.h" #include "host_ir/lowering.h" #include "host_ir/passes.h" #include "instrumentation.h" #include "ir/base_nodes.h" #include "preseg_passes/pre_segmenter.h" #include "runtime/executor.h" #include "runtime/executor_dispatch.h" #include "runtime/fusion_cache_utils.h" #include "scheduler/heuristic.h" #include "serde/fusion_cache_generated.h" #include "type.h"

Test failures

• (Medium, 3) Thunder thunderfx path returns scalar instead of vector in higher-order inplace alias update (thunder.tests.test_update_aliases)

  Test Name	A100	GB200	H100	Source
  thunder.tests.test_update_aliases.test_higher_order_inplace_alias_update_nvfuser_cuda_thunder.dtypes.float32	❌	❌	❌

• (Medium, 3) ModuleNotFoundError for 'nvfuser' in tests.python.direct.test_import

  Test Name	A100	GB200	H100	Source
  tests.python.direct.test_import.test_import_conflict_direct_then_nvfuser	❌	❌	❌

[greptile-apps]

Greptile Overview

Greptile Summary

This PR removes the legacy nvfuser Python module and its associated pybind11 bindings. The version is bumped to 0.2.36.

Major changes:

• Removed entire python/python_frontend/ directory containing legacy pybind11 bindings
• Removed nvfuser Python package (649 lines from __init__.py, utilities, contrib modules)
• Cleaned up CMakeLists.txt by removing nvfuser target and associated build configuration
• Removed csrc/serde/fusion_record.cpp (952 lines) and associated headers
• Removed Python frontend debug options from csrc/options.h
• Updated python/utils.py to only build nvfuser_direct extension

Issues found:

• Several files still import from the removed nvfuser module, which will cause import errors
• tests/python/utils/utils.py imports from nvfuser
• Documentation files in doc/dev/python_scheduling/ (6 files) still import from nvfuser

These files need to be updated to import from nvfuser_direct instead to maintain functionality.

Confidence Score: 3/5

• This PR has breaking import errors that need to be fixed before merging
• The PR successfully removes the legacy bindings infrastructure, but leaves broken imports in test utilities and documentation that will cause runtime failures. The core changes are clean, but the incomplete migration of imports prevents this from being merge-ready.
• tests/python/utils/utils.py and all files in doc/dev/python_scheduling/ need their imports updated from nvfuser to nvfuser_direct

Important Files Changed

Filename	Overview
tests/python/direct/test_import.py	Removed test for nvfuser import conflict
python/utils.py	Removed nvfuser._C extension from build configuration
CMakeLists.txt	Removed nvfuser Python bindings and legacy pybind11 targets
tests/python/utils/utils.py	Still imports from nvfuser module which no longer exists
doc/dev/python_scheduling/autotune_utils.py	Still imports from nvfuser module which no longer exists

Sequence Diagram

sequenceDiagram
    participant User
    participant Python
    participant nvfuser_direct
    participant libnvfuser_codegen
    
    Note over nvfuser: Legacy module REMOVED
    
    User->>Python: import nvfuser_direct
    Python->>nvfuser_direct: Load _C_DIRECT extension
    nvfuser_direct->>libnvfuser_codegen: Link to shared library
    libnvfuser_codegen-->>nvfuser_direct: Core codegen functionality
    nvfuser_direct-->>Python: Module loaded
    Python-->>User: Success
    
    Note over User,Python: Old workflow (NOW BROKEN):
    User->>Python: import nvfuser
    Python-->>User: ImportError: No module named 'nvfuser'

Loading

[greptile-apps]

_{7 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Additional Comments (2)

python/utils.py
stale reference to removed python_frontend directory

---

tools/env-config/env_options.yaml
stale entries for removed debug options python_definition_segments and python_frontend_debug should be removed

[rdspring1]

!build

[greptile-apps]

_{8 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Additional Comments (1)

python/utils.py
outdated comment references nvfuser._C which was removed in this PR

    # "nvfuser_codegen" into "./nvfuser_direct/lib", and the former will be "nvfuser_direct._C_DIRECT".

[rdspring1]

!test

[greptile-apps]

_{8 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[rdspring1]

!test

[greptile-apps]

_{5 files reviewed, 7 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Additional Comments (7)

tests/python/utils/utils.py
The nvfuser module has been removed in this PR. Update to import from nvfuser_direct instead:

from nvfuser_direct import FusionCache, FusionDefinition, DataType, clone, Tensor

---

doc/dev/python_scheduling/autotune_utils.py
The nvfuser module has been removed. Update to import from nvfuser_direct:

from nvfuser_direct import FusionCache, FusionDefinition

---

doc/dev/python_scheduling/autotune_pointwise.py
Update to use nvfuser_direct instead of removed nvfuser module:

from nvfuser_direct import FusionDefinition, SchedulerType, DataType

---

doc/dev/python_scheduling/autotune_inner_reduction.py
Update to use nvfuser_direct:

from nvfuser_direct import FusionDefinition, SchedulerType, DataType, ParallelType

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}

---

doc/dev/python_scheduling/autotune_persistent.py
Update imports to use nvfuser_direct:

from nvfuser_direct import FusionCache, FusionDefinition, SchedulerType, DataType
from nvfuser_direct.pytorch_utils import torch_dtype_to_nvfuser_dtype

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}

---

doc/dev/python_scheduling/autotune_matmul.py
Update to use nvfuser_direct:

from nvfuser_direct import FusionDefinition, SchedulerType

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}

---

doc/dev/python_scheduling/profile_matmul.py
Update imports to use nvfuser_direct instead of removed nvfuser module