BUG: Pymc receiving torch tensors causes sampling failures

[Ch0ronomato]

Describe the issue:

If you allow the torch backend for pytensor, and try to sample a simple model, you will receive dtype errors during mcmc initialization.

Reproducable code example:

import pymc as pm
import pytensor as pt

pt.config.mode = "PYTORCH" # allow the backend in configdefaults.py

with pm.Model() as m:
    x = pm.Beta(name="test", alpha=0.5, beta=0.5)
    pm.sample(3)

Error message:

<details>
(pytensor-dev)  ch0ronomato@macbook-pro  ~/dev/pytensor   torch_probs_goofin ●  python test_torch_pymc.py                    
Only 3 samples per chain. Reliable r-hat and ESS diagnostics require longer chains for accurate estimate.
Auto-assigning NUTS sampler...
Initializing NUTS using jitter+adapt_diag...
Traceback (most recent call last):
  File "/Users/ch0ronomato/dev/pytensor/test_torch_pymc.py", line 8, in <module>
    pm.sample(3)
  File "/opt/anaconda3/envs/pytensor-dev/lib/python3.11/site-packages/pymc/sampling/mcmc.py", line 789, in sample
    run, traces = init_traces(
                  ^^^^^^^^^^^^
  File "/opt/anaconda3/envs/pytensor-dev/lib/python3.11/site-packages/pymc/backends/__init__.py", line 140, in init_traces
    traces = [
             ^
  File "/opt/anaconda3/envs/pytensor-dev/lib/python3.11/site-packages/pymc/backends/__init__.py", line 141, in <listcomp>
    _init_trace(
  File "/opt/anaconda3/envs/pytensor-dev/lib/python3.11/site-packages/pymc/backends/__init__.py", line 115, in _init_trace
    strace.setup(expected_length, chain_number, stats_dtypes)
  File "/opt/anaconda3/envs/pytensor-dev/lib/python3.11/site-packages/pymc/backends/ndarray.py", line 79, in setup
    self.samples[varname] = np.zeros((draws, *shape), dtype=self.var_dtypes[varname])
                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: Cannot interpret 'torch.float64' as a data type
</details>

PyTensor version information:

floatX ({'float64', 'float32', 'float16'}) Doc: Default floating-point precision for python casts.

Note: float16 support is experimental, use at your own risk.
Value: float64

warn_float64 ({'pdb', 'raise', 'ignore', 'warn'})
Doc: Do an action when a tensor variable with float64 dtype is created.
Value: ignore

pickle_test_value (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x10fb63e50>>)
Doc: Dump test values while pickling model. If True, test values will be dumped with model.
Value: True

cast_policy ({'numpy+floatX', 'custom'})
Doc: Rules for implicit type casting
Value: custom

device (cpu)
Doc: Default device for computations. only cpu is supported for now
Value: cpu

conv__assert_shape (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x10f9492d0>>)
Doc: If True, AbstractConv* ops will verify that user-provided shapes match the runtime shapes (debugging option, may slow down compilation)
Value: False

print_global_stats (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x10fcb6850>>)
Doc: Print some global statistics (time spent) at the end
Value: False

unpickle_function (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x10fc191d0>>)
Doc: Replace unpickled PyTensor functions with None. This is useful to unpickle old graphs that pickled them when it shouldn't
Value: True

<pytensor.configparser.ConfigParam object at 0x1100e0110>
Doc: Default compilation mode
Value: Mode

cxx (<class 'str'>)
Doc: The C++ compiler to use. Currently only g++ is supported, but supporting additional compilers should not be too difficult. If it is empty, no C++ code is compiled.
Value: /opt/anaconda3/envs/pytensor-dev/bin/clang++

linker ({'vm_nogc', 'cvm', 'vm', 'cvm_nogc', 'c|py', 'c|py_nogc', 'c', 'py'})
Doc: Default linker used if the pytensor flags mode is Mode
Value: cvm

allow_gc (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x10fca0450>>)
Doc: Do we default to delete intermediate results during PyTensor function calls? Doing so lowers the memory requirement, but asks that we reallocate memory at the next function call. This is implemented for the default linker, but may not work for all linkers.
Value: True

optimizer ({'o1', 'o2', 'fast_run', 'o4', 'o3', 'None', 'unsafe', 'fast_compile', 'merge'})
Doc: Default optimizer. If not None, will use this optimizer with the Mode
Value: o4

optimizer_verbose (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x10f94bf90>>)
Doc: If True, we print all optimization being applied
Value: False

on_opt_error ({'pdb', 'raise', 'ignore', 'warn'})
Doc: What to do when an optimization crashes: warn and skip it, raise the exception, or fall into the pdb debugger.
Value: warn

nocleanup (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x110b6c3d0>>)
Doc: Suppress the deletion of code files that did not compile cleanly
Value: False

on_unused_input ({'raise', 'ignore', 'warn'})
Doc: What to do if a variable in the 'inputs' list of pytensor.function() is not used in the graph.
Value: raise

gcc__cxxflags (<class 'str'>)
Doc: Extra compiler flags for gcc
Value: -Wno-c++11-narrowing

cmodule__warn_no_version (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x110b6d690>>)
Doc: If True, will print a warning when compiling one or more Op with C code that can't be cached because there is no c_code_cache_version() function associated to at least one of those Ops.
Value: False

cmodule__remove_gxx_opt (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x110027050>>)
Doc: If True, will remove the -O* parameter passed to g++.This is useful to debug in gdb modules compiled by PyTensor.The parameter -g is passed by default to g++
Value: False

cmodule__compilation_warning (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x110b6c310>>)
Doc: If True, will print compilation warnings.
Value: False

cmodule__preload_cache (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x110b6c290>>)
Doc: If set to True, will preload the C module cache at import time
Value: False

cmodule__age_thresh_use (<class 'int'>)
Doc: In seconds. The time after which PyTensor won't reuse a compile c module.
Value: 2073600

cmodule__debug (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x10fc19690>>)
Doc: If True, define a DEBUG macro (if not exists) for any compiled C code.
Value: False

compile__wait (<class 'int'>)
Doc: Time to wait before retrying to acquire the compile lock.
Value: 5

compile__timeout (<class 'int'>)
Doc: In seconds, time that a process will wait before deciding to
override an existing lock. An override only happens when the existing
lock is held by the same owner and has not been 'refreshed' by this
owner for more than this period. Refreshes are done every half timeout
period for running processes.
Value: 120

tensor__cmp_sloppy (<class 'int'>)
Doc: Relax pytensor.tensor.math._allclose (0) not at all, (1) a bit, (2) more
Value: 0

lib__amdlibm (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x11010bc90>>)
Doc: Use amd's amdlibm numerical library
Value: False

tensor__insert_inplace_optimizer_validate_nb (<class 'int'>)
Doc: -1: auto, if graph have less then 500 nodes 1, else 10
Value: -1

traceback__limit (<class 'int'>)
Doc: The number of stack to trace. -1 mean all.
Value: 8

traceback__compile_limit (<class 'int'>)
Doc: The number of stack to trace to keep during compilation. -1 mean all. If greater then 0, will also make us save PyTensor internal stack trace.
Value: 0

warn__ignore_bug_before ({'0.9', '0.4.1', '0.8', '0.4', '0.3', '1.0.2', '1.0.1', '0.5', '1.0.3', '0.10', '1.0.4', 'all', '0.8.1', 'None', '1.0', '0.8.2', '0.6', '1.0.5', '0.7'})
Doc: If 'None', we warn about all PyTensor bugs found by default. If 'all', we don't warn about PyTensor bugs found by default. If a version, we print only the warnings relative to PyTensor bugs found after that version. Warning for specific bugs can be configured with specific [warn] flags.
Value: 0.9

exception_verbosity ({'low', 'high'})
Doc: If 'low', the text of exceptions will generally refer to apply nodes with short names such as Elemwise{add_no_inplace}. If 'high', some exceptions will also refer to apply nodes with long descriptions like:
A. Elemwise{add_no_inplace}
B. log_likelihood_v_given_h
C. log_likelihood_h
Value: low

print_test_value (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x11010aa10>>)
Doc: If 'True', the eval of an PyTensor variable will return its test_value when this is available. This has the practical consequence that, e.g., in debugging my_var will print the same as my_var.tag.test_value when a test value is defined.
Value: False

compute_test_value ({'off', 'raise', 'warn', 'pdb', 'ignore'})
Doc: If 'True', PyTensor will run each op at graph build time, using Constants, SharedVariables and the tag 'test_value' as inputs to the function. This helps the user track down problems in the graph before it gets optimized.
Value: off

compute_test_value_opt ({'off', 'raise', 'warn', 'pdb', 'ignore'})
Doc: For debugging PyTensor optimization only. Same as compute_test_value, but is used during PyTensor optimization
Value: off

check_input (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x110b6c050>>)
Doc: Specify if types should check their input in their C code. It can be used to speed up compilation, reduce overhead (particularly for scalars) and reduce the number of generated C files.
Value: True

NanGuardMode__nan_is_error (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x110b6c0d0>>)
Doc: Default value for nan_is_error
Value: True

NanGuardMode__inf_is_error (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x1100e2950>>)
Doc: Default value for inf_is_error
Value: True

NanGuardMode__big_is_error (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x110b6d450>>)
Doc: Default value for big_is_error
Value: True

NanGuardMode__action ({'pdb', 'raise', 'warn'})
Doc: What NanGuardMode does when it finds a problem
Value: raise

DebugMode__patience (<class 'int'>)
Doc: Optimize graph this many times to detect inconsistency
Value: 10

DebugMode__check_c (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x10fedf390>>)
Doc: Run C implementations where possible
Value: True

DebugMode__check_py (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x1100e3d10>>)
Doc: Run Python implementations where possible
Value: True

DebugMode__check_finite (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x110b6da90>>)
Doc: True -> complain about NaN/Inf results
Value: True

DebugMode__check_strides (<class 'int'>)
Doc: Check that Python- and C-produced ndarrays have same strides. On difference: (0) - ignore, (1) warn, or (2) raise error
Value: 0

DebugMode__warn_input_not_reused (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x10fb73390>>)
Doc: Generate a warning when destroy_map or view_map says that an op works inplace, but the op did not reuse the input for its output.
Value: True

DebugMode__check_preallocated_output (<class 'str'>)
Doc: Test thunks with pre-allocated memory as output storage. This is a list of strings separated by ":". Valid values are: "initial" (initial storage in storage map, happens with Scan),"previous" (previously-returned memory), "c_contiguous", "f_contiguous", "strided" (positive and negative strides), "wrong_size" (larger and smaller dimensions), and "ALL" (all of the above).
Value:

DebugMode__check_preallocated_output_ndim (<class 'int'>)
Doc: When testing with "strided" preallocated output memory, test all combinations of strides over that number of (inner-most) dimensions. You may want to reduce that number to reduce memory or time usage, but it is advised to keep a minimum of 2.
Value: 4

profiling__time_thunks (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x1100e2a10>>)
Doc: Time individual thunks when profiling
Value: True

profiling__n_apply (<class 'int'>)
Doc: Number of Apply instances to print by default
Value: 20

profiling__n_ops (<class 'int'>)
Doc: Number of Ops to print by default
Value: 20

profiling__output_line_width (<class 'int'>)
Doc: Max line width for the profiling output
Value: 512

profiling__min_memory_size (<class 'int'>)
Doc: For the memory profile, do not print Apply nodes if the size
of their outputs (in bytes) is lower than this threshold
Value: 1024

profiling__min_peak_memory (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x1100e29d0>>)
Doc: The min peak memory usage of the order
Value: False

profiling__destination (<class 'str'>)
Doc: File destination of the profiling output
Value: stderr

profiling__debugprint (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x10fc1a990>>)
Doc: Do a debugprint of the profiled functions
Value: False

profiling__ignore_first_call (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x110b6dc90>>)
Doc: Do we ignore the first call of an PyTensor function.
Value: False

on_shape_error ({'raise', 'warn'})
Doc: warn: print a warning and use the default value. raise: raise an error
Value: warn

openmp (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x1100e3d90>>)
Doc: Allow (or not) parallel computation on the CPU with OpenMP. This is the default value used when creating an Op that supports OpenMP parallelization. It is preferable to define it via the PyTensor configuration file ~/.pytensorrc or with the environment variable PYTENSOR_FLAGS. Parallelization is only done for some operations that implement it, and even for operations that implement parallelism, each operation is free to respect this flag or not. You can control the number of threads used with the environment variable OMP_NUM_THREADS. If it is set to 1, we disable openmp in PyTensor by default.
Value: False

openmp_elemwise_minsize (<class 'int'>)
Doc: If OpenMP is enabled, this is the minimum size of vectors for which the openmp parallelization is enabled in element wise ops.
Value: 200000

optimizer_excluding (<class 'str'>)
Doc: When using the default mode, we will remove optimizer with these tags. Separate tags with ':'.
Value:

optimizer_including (<class 'str'>)
Doc: When using the default mode, we will add optimizer with these tags. Separate tags with ':'.
Value:

optimizer_requiring (<class 'str'>)
Doc: When using the default mode, we will require optimizer with these tags. Separate tags with ':'.
Value:

optdb__position_cutoff (<class 'float'>)
Doc: Where to stop earlier during optimization. It represent the position of the optimizer where to stop.
Value: inf

optdb__max_use_ratio (<class 'float'>)
Doc: A ratio that prevent infinite loop in EquilibriumGraphRewriter.
Value: 8.0

cycle_detection ({'regular', 'fast'})
Doc: If cycle_detection is set to regular, most inplaces are allowed,but it is slower. If cycle_detection is set to faster, less inplacesare allowed, but it makes the compilation faster.The interaction of which one give the lower peak memory usage iscomplicated and not predictable, so if you are close to the peakmemory usage, triyng both could give you a small gain.
Value: regular

check_stack_trace ({'raise', 'log', 'off', 'warn'})
Doc: A flag for checking the stack trace during the optimization process. default (off): does not check the stack trace of any optimization log: inserts a dummy stack trace that identifies the optimizationthat inserted the variable that had an empty stack trace.warn: prints a warning if a stack trace is missing and also a dummystack trace is inserted that indicates which optimization insertedthe variable that had an empty stack trace.raise: raises an exception if a stack trace is missing
Value: off

metaopt__verbose (<class 'int'>)
Doc: 0 for silent, 1 for only warnings, 2 for full output withtimings and selected implementation
Value: 0

unittests__rseed (<class 'str'>)
Doc: Seed to use for randomized unit tests. Special value 'random' means using a seed of None.
Value: 666

warn__round (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x1100e0150>>)
Doc: Warn when using tensor.round with the default mode. Round changed its default from half_away_from_zero to half_to_even to have the same default as NumPy.
Value: False

profile (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x10f8071d0>>)
Doc: If VM should collect profile information
Value: False

profile_optimizer (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x110b6df10>>)
Doc: If VM should collect optimizer profile information
Value: False

profile_memory (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x10fbdf290>>)
Doc: If VM should collect memory profile information and print it
Value: False

<pytensor.configparser.ConfigParam object at 0x110b6e050>
Doc: Useful only for the VM Linkers. When lazy is None, auto detect if lazy evaluation is needed and use the appropriate version. If the C loop isn't being used and lazy is True, use the Stack VM; otherwise, use the Loop VM.
Value: None

numba__vectorize_target ({'cpu', 'parallel', 'cuda'})
Doc: Default target for numba.vectorize.
Value: cpu

numba__fastmath (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x10fb63cd0>>)
Doc: If True, use Numba's fastmath mode.
Value: True

numba__cache (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x110b6e250>>)
Doc: If True, use Numba's file based caching.
Value: True

compiledir_format (<class 'str'>)
Doc: Format string for platform-dependent compiled module subdirectory
(relative to base_compiledir). Available keys: device, gxx_version,
hostname, numpy_version, platform, processor, pytensor_version,
python_bitwidth, python_int_bitwidth, python_version, short_platform.
Defaults to compiledir_%(short_platform)s-%(processor)s-
%(python_version)s-%(python_bitwidth)s.
Value: compiledir_%(short_platform)s-%(processor)s-%(python_version)s-%(python_bitwidth)s

<pytensor.configparser.ConfigParam object at 0x1100ed2d0>
Doc: platform-independent root directory for compiled modules
Value: /Users/ch0ronomato/.pytensor

<pytensor.configparser.ConfigParam object at 0x10fb63e10>
Doc: platform-dependent cache directory for compiled modules
Value: /Users/ch0ronomato/.pytensor/compiledir_macOS-14.5-x86_64-i386-64bit-i386-3.11.9-64

blas__ldflags (<class 'str'>)
Doc: lib[s] to include for [Fortran] level-3 blas implementation
Value: -L/opt/anaconda3/envs/pytensor-dev/lib -llapack -lblas -lcblas -lm -Wl,-rpath,/opt/anaconda3/envs/pytensor-dev/lib

blas__check_openmp (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x1135c7410>>)
Doc: Check for openmp library conflict.
WARNING: Setting this to False leaves you open to wrong results in blas-related operations.
Value: True

scan__allow_gc (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x125fe3690>>)
Doc: Allow/disallow gc inside of Scan (default: False)
Value: False

scan__allow_output_prealloc (<bound method BoolParam._apply of <pytensor.configparser.BoolParam object at 0x125f609d0>>)
Doc: Allow/disallow memory preallocation for outputs inside of scan (default: True)
Value: True

Context for the issue:

I started goofing around to see if pymc just works with torch while I do some of the probability distributions.

[ricardoV94]

The problem is the trace is using the initial point, from model.initial_point() to set the buffers and using value.shape and value.dtype, which when the initial_point is compiled with PyTorch backend ends up being torch.float64 that numpy does not recognize.