…nding

- Add setFloat8_/getFloat8_ function pointer members to Variable.h
- Add public setFloat8/getFloat8 accessor declarations to Variable.h
- Add case rim::Float8 to all four switch blocks in Variable.cpp
- Implement setFloat8/getFloat8 C++ accessor wrappers in Variable.cpp
- Expose Float8 constant to Python in module.cpp

- Add Float8(Model) with E4M3 encoding: modelId=rim.Float8, ndType=float32
- Manual bit manipulation toBytes/fromBytes (no struct.pack format for 8-bit float)
- NaN encodes as 0x7F, infinity clamps to max finite (sign|0x7E)
- minValue=-448.0, maxValue=448.0
- Float8BE subclass with endianness='big'

- Create tests/core/test_float8.py with 18 tests covering metadata,
  boundary values round-trip, NaN encoding (0x7F), no-infinity clamping,
  known bit patterns, and RemoteVariable integration
- Add Float8Var (offset=0x30, bitSize=8) to ModelVariableDevice in
  test_model_variables.py and exercise it in round-trip test
- All 22 tests pass

- Create docs/src/api/python/pyrogue/float8.rst with autoclass directives
- Add float8 to Models toctree in index.rst (after float16)
- Add Float8/Float8BE rows to Built-In Model Types table in model.rst
- Update floating-point family list to include Float8 and Float8BE

…thods

- Add BFloat16 = 0x0B constant to Constants.h after Float8
- Add setBFloat16Py/getBFloat16Py/setBFloat16/getBFloat16 declarations to Block.h
- Add floatToBFloat16/bfloat16ToFloat converter functions (upper 16 bits of float32)
- Implement all four BFloat16 Block methods using NPY_FLOAT (no native numpy dtype)

…binding

- Add setBFloat16_/getBFloat16_ function pointer members to Variable.h
- Add setBFloat16/getBFloat16 public accessor declarations to Variable.h
- Add four case rim::BFloat16 dispatch blocks in Variable.cpp constructor
- Add setBFloat16/getBFloat16 C++ accessor wrapper methods in Variable.cpp
- Initialize setBFloat16_/getBFloat16_ to NULL in constructor
- Expose rim.BFloat16 = 0x0B constant via module.cpp Python binding

- BFloat16(Model): 1s/8e/7m format, upper 16 bits of float32 bit pattern
- modelId = rim.BFloat16, ndType = np.float32, bitSize must be 16
- toBytes/fromBytes use integer bit manipulation (no struct format code for BF16)
- Supports infinity and NaN (unlike Float8 which clamps)
- BFloat16BE subclass with endianness = 'big'

- Create tests/core/test_bfloat16.py with 17 tests covering:
  metadata, wrong bitsize rejection, instance caching, boundary value
  round-trips, NaN, infinity, known bit patterns, BE endianness,
  and remote variable integration through C++ Block layer
- Update test_model_variables.py: add BFloat16Var at offset=0x32
  and include it in round-trip test parametrization

…nding

- Add setTensorFloat32_/getTensorFloat32_ function pointers and public accessors to Variable.h
- Add NULL init, 4 switch-case dispatches, and C++ wrapper methods to Variable.cpp
- Expose TensorFloat32 constant to Python via module.cpp
- rim.TensorFloat32 == 12 verified from Python

- TensorFloat32 class with 4-byte storage, 0xFFFFE000 mask, rim.TensorFloat32 modelId
- TensorFloat32BE big-endian variant
- bitSize=32 assertion, ndType=np.float32, min/max ~3.40e38

…tion

- 17 tests: metadata, wrong bitsize, cache, boundary round-trip, NaN, infinity, fromString, bit patterns, BE endianness, remote variable round-trip, remote variable boundary values
- TF32Var added to model variables integration device at offset=0x34, bitSize=32
- All 35 existing BFloat16/Float8 tests pass with no regressions

- New tensorfloat32.rst with autoclass directives for TensorFloat32 and TensorFloat32BE
- Added tensorfloat32 to Models toctree in pyrogue index.rst
- Added TF32 rows to Built-In Model Types table and floating-point family list in model.rst

… get/set methods

- Add Float6 = 0x0D constant in Constants.h
- Add setFloat6/getFloat6/setFloat6Py/getFloat6Py declarations in Block.h
- Implement floatToFloat6/float6ToFloat converters with E3M2 semantics (1s/3e/2m, bias=3, no NaN/Inf)
- Implement all four Block methods with NPY_FLOAT numpy array support

…nding

- Add setFloat6_/getFloat6_ function pointer members in Variable.h
- Add public setFloat6/getFloat6 accessor declarations in Variable.h
- Add 4 case rim::Float6 dispatch blocks in Variable.cpp
- Add NULL initialization for Float6 function pointers
- Add C++ accessor wrapper implementations
- Expose Float6 constant to Python in module.cpp

- Float6 E3M2 model with toBytes/fromBytes bit manipulation
- NaN/Inf inputs clamp to max finite (+/-28.0)
- Float6BE big-endian variant
- bitSize=8, ndType=float32, modelId=rim.Float6

- Create test_float6.py with metadata, boundary, bit pattern, NaN/Inf, remote variable tests
- Add Float6Var to test_model_variables.py integration test
- Fix float6ToFloat C++ subnormal decode using direct formula instead of normalization loop

- Create float6.rst with autoclass directives for Float6/Float6BE
- Add float6 to API index toctree
- Add Float6/Float6BE rows to model type table and family list

… methods

- Add Float4 = 0x0E constant in Constants.h
- Add setFloat4/getFloat4/setFloat4Py/getFloat4Py declarations in Block.h
- Implement floatToFloat4/float4ToFloat converters with E2M1 semantics (1s/2e/1m, bias=1, no NaN/Inf)
- Implement Block set/get methods for Float4 with NPY_FLOAT numpy array support

…nding

- Add setFloat4_/getFloat4_ function pointer members in Variable.h
- Add public setFloat4/getFloat4 accessor declarations in Variable.h
- Add 4 case rim::Float4: blocks in Variable.cpp switch statements
- Add NULL initialization for Float4 function pointers
- Add C++ accessor wrapper implementations for Float4
- Expose Float4 constant to Python in module.cpp

- Float4 E2M1 model with manual bit manipulation encode/decode
- NaN/Inf inputs clamp to max finite (+/-6.0), no special encodings
- Float4BE big-endian variant
- ndType=float32, bitSize assertion enforces 8-bit storage

…type

The normalization loop in float8ToFloat checked bit 2 (0x04) for the
implicit leading 1, but Float8's 3-bit stored mantissa has its leading
bit at position 3 (0x08). Stopping one shift early produced a 2x error
for every subnormal value (e.g. 0x01 decoded to 2^-8 instead of 2^-9).

Fix by checking bit 3 (0x08) and stripping with 0x07 (not 0x03), and
use int32_t for the working exponent so the counter can go negative
without wrapping. Add a RemoteVariable subnormal test (2^-9) to cover
the C++ Block conversion path.

The pre-release merge brought the wide signed Fixed32/40/64 test
variables into test_model_variables.py alongside the Float16/8/BFloat16/
TF32/Float6/Float4 variables already on this branch. Both sets used
offsets 0x28/0x30/0x38, so Block::addVariables rejected the device with
a bit-overlap GeneralError and every test in the file failed at
construction time. Move the Fixed* entries into the free region after
UFixedListVar (0x68/0x70/0x78), which is well within the 0x4000 emulated
memory window.

…del.rst

Convert the new Float8/BFloat16/TensorFloat32/Float6/Float4 rows (and BE
variants) in the Built-In Model Types table to use :ref: cross-references
consistent with every other row. Correct the Bit Size column for Float6
and Float4 from 8-bits (the storage size) to 6-bits and 4-bits (the format
width). Widen columns so BFloat16/TF32 notes and the tensorfloat32be ref
fit within their grid cells.