[REFACTOR][DataType] Phase out target custom datatype support (#19760)

## Summary

The in-tree target custom datatype path adds maintenance surface while
current development focuses on core datatypes. This PR phases out the
built-in registry/lowering implementation and keeps the core dtype
behavior intact.

- Remove the target/datatype implementation, BYODT posit build option,
and related Python helpers
- Remove the custom datatype lowering pass from TIRX and S-TIR
finalization pipelines
- Simplify remaining TIRX dtype handling back to built-in/core datatypes

Author: tqchen

CMakeLists.txt

@@ -87,7 +87,6 @@ tvm_option(USE_CCACHE "Use ccache if found when invoking compiler" AUTO)
 # 3rdparty libraries
 tvm_option(COMPILER_RT_PATH "Path to COMPILER-RT" "3rdparty/compiler-rt")
 # Contrib library options
-tvm_option(USE_BYODT_POSIT "Build with BYODT software emulated posit custom datatype" OFF)
 tvm_option(USE_BLAS "The blas library to be linked" none)
 tvm_option(USE_AMX "Enable Intel AMX" OFF)
 tvm_option(USE_MKL "MKL root path when use MKL blas" OFF)
@@ -356,10 +355,6 @@ tvm_file_glob(GLOB CODEGEN_SRCS
 
 list(APPEND COMPILER_SRCS ${CODEGEN_SRCS})
 
-tvm_file_glob(GLOB DATATYPE_SRCS src/target/datatype/*.cc)
-list(APPEND COMPILER_SRCS ${DATATYPE_SRCS})
-list(APPEND COMPILER_SRCS "src/target/datatype/myfloat/myfloat.cc")
-
 tvm_file_glob(GLOB RUNTIME_SRCS
   src/runtime/*.cc
   src/runtime/vm/*.cc
@@ -464,7 +459,6 @@ include(cmake/modules/contrib/DNNL.cmake)
 include(cmake/modules/contrib/AMX.cmake)
 include(cmake/modules/contrib/CUTLASS.cmake)
 include(cmake/modules/contrib/Random.cmake)
-include(cmake/modules/contrib/Posit.cmake)
 include(cmake/modules/contrib/Sort.cmake)
 include(cmake/modules/contrib/CoreML.cmake)
 include(cmake/modules/contrib/TensorRT.cmake)

cmake/modules/contrib/Posit.cmake

@@ -63,10 +63,6 @@ RUN bash /install/ubuntu_install_dnnl.sh
 COPY install/ubuntu_install_xgboost.sh /install/ubuntu_install_xgboost.sh
 RUN bash /install/ubuntu_install_xgboost.sh
 
-# BYODT deps
-COPY install/ubuntu_install_universal.sh /install/ubuntu_install_universal.sh
-RUN bash /install/ubuntu_install_universal.sh
-
 # TensorFlow deps
 COPY install/ubuntu_install_tensorflow.sh /install/ubuntu_install_tensorflow.sh
 RUN bash /install/ubuntu_install_tensorflow.sh

docker/Dockerfile.ci_cpu

@@ -115,10 +115,6 @@ RUN bash /install/ubuntu_install_vulkan.sh
 COPY install/ubuntu_install_xgboost.sh /install/ubuntu_install_xgboost.sh
 RUN bash /install/ubuntu_install_xgboost.sh
 
-# BYODT deps
-COPY install/ubuntu_install_universal.sh /install/ubuntu_install_universal.sh
-RUN bash /install/ubuntu_install_universal.sh
-
 # sccache
 COPY install/ubuntu_install_sccache.sh /install/ubuntu_install_sccache.sh
 RUN bash /install/ubuntu_install_sccache.sh

docker/Dockerfile.ci_gpu

@@ -998,13 +998,6 @@ inline PrimExpr MakeConstScalar(DataType t, ValueType value, Span span = Span())
   }
   if (t.is_float() || t.is_bfloat16() || t.is_float8() || t.is_float6() || t.is_float4())
     return FloatImm(t, static_cast<double>(value), span);
-  // For now, we store const scalar values of custom datatypes within doubles; later, during the
-  // datatypes lowering pass, we will lower the value to its true representation in the format
-  // specified by the datatype.
-  // TODO(gus) when do we need to start worrying about doubles not being precise enough?
-  if (static_cast<uint8_t>(t.code()) >= static_cast<uint8_t>(DataType::kCustomBegin)) {
-    return FloatImm(t, static_cast<double>(value), span);
-  }
   TVM_FFI_THROW(InternalError) << "cannot make const for type " << t;
   throw;
 }

include/tvm/tirx/op.h

@@ -153,15 +153,6 @@ TVM_DLL Pass MakePackedAPI();
  */
 TVM_DLL Pass RemapThreadAxis(ffi::Map<ffi::String, IterVar> axis_map);
 
-/*!
- * \brief Lower custom datatypes.
- *
- * See tvm::datatypes::Registry for more information on adding custom datatypes.
- *
- * \return The pass.
- */
-TVM_DLL Pass LowerCustomDatatypes();
-
 /*!
  * \brief Annotate, split, and lower host/device functions.
  *

include/tvm/tirx/transform.h

@@ -125,7 +125,6 @@ def finalize_host_passes():  # pylint: disable=unused-argument
     """The default finalization passes for TIR backend."""
     host_pass_list = [
         tirx.transform.LowerTVMBuiltin(),
-        tirx.transform.LowerCustomDatatypes(),
         tirx.transform.LowerIntrin(),
     ]
     return tvm.ir.transform.Sequential(host_pass_list)
@@ -136,7 +135,6 @@ def finalize_device_passes():  # pylint: disable=unused-argument
     device_pass_list = [
         tirx.transform.LowerWarpMemory(),
         tirx.transform.StmtSimplify(),
-        tirx.transform.LowerCustomDatatypes(),
         tirx.transform.LowerIntrin(),
     ]
     return tvm.ir.transform.Sequential(device_pass_list)

python/tvm/s_tir/pipeline.py

@@ -34,6 +34,5 @@
 from .target import Target, TargetKind
 from .virtual_device import VirtualDevice
 from .tag import list_tags, register_tag
-from . import datatype
 from . import codegen
 from . import tag_registry  # registers tags on import