warp-lang 1.6.2__py3-none-macosx_10_13_universal2.whl → 1.7.0__py3-none-macosx_10_13_universal2.whl

warp/context.py

@@ -17,7 +17,6 @@ from __future__ import annotations
 
 import ast
 import ctypes
-import errno
 import functools
 import hashlib
 import inspect
@@ -28,13 +27,27 @@ import operator
 import os
 import platform
 import sys
-import time
 import types
 import typing
 import weakref
 from copy import copy as shallowcopy
 from pathlib import Path
-from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple, Union
+from typing import (
+    Any,
+    Callable,
+    Dict,
+    List,
+    Literal,
+    Mapping,
+    Optional,
+    Sequence,
+    Set,
+    Tuple,
+    TypeVar,
+    Union,
+    get_args,
+    get_origin,
+)
 
 import numpy as np
 
@@ -42,7 +55,7 @@ import warp
 import warp.build
 import warp.codegen
 import warp.config
-from warp.types import launch_bounds_t
+from warp.types import Array, launch_bounds_t
 
 # represents either a built-in or user-defined function
 
@@ -71,10 +84,10 @@ def get_function_args(func):
 complex_type_hints = (Any, Callable, Tuple)
 sequence_types = (list, tuple)
 
-function_key_counts = {}
+function_key_counts: Dict[str, int] = {}
 
 
-def generate_unique_function_identifier(key):
+def generate_unique_function_identifier(key: str) -> str:
     # Generate unique identifiers for user-defined functions in native code.
     # - Prevents conflicts when a function is redefined and old versions are still in use.
     # - Prevents conflicts between multiple closures returned from the same function.
@@ -107,40 +120,40 @@ def generate_unique_function_identifier(key):
 class Function:
     def __init__(
         self,
-        func,
-        key,
-        namespace,
-        input_types=None,
-        value_type=None,
-        value_func=None,
-        export_func=None,
-        dispatch_func=None,
-        lto_dispatch_func=None,
-        module=None,
-        variadic=False,
-        initializer_list_func=None,
-        export=False,
-        doc="",
-        group="",
-        hidden=False,
-        skip_replay=False,
-        missing_grad=False,
-        generic=False,
-        native_func=None,
-        defaults=None,
-        custom_replay_func=None,
-        native_snippet=None,
-        adj_native_snippet=None,
-        replay_snippet=None,
-        skip_forward_codegen=False,
-        skip_reverse_codegen=False,
-        custom_reverse_num_input_args=-1,
-        custom_reverse_mode=False,
-        overloaded_annotations=None,
-        code_transformers=None,
-        skip_adding_overload=False,
-        require_original_output_arg=False,
-        scope_locals=None,  # the locals() where the function is defined, used for overload management
+        func: Optional[Callable],
+        key: str,
+        namespace: str,
+        input_types: Optional[Dict[str, Union[type, TypeVar]]] = None,
+        value_type: Optional[type] = None,
+        value_func: Optional[Callable[[Mapping[str, type], Mapping[str, Any]], type]] = None,
+        export_func: Optional[Callable[[Dict[str, type]], Dict[str, type]]] = None,
+        dispatch_func: Optional[Callable] = None,
+        lto_dispatch_func: Optional[Callable] = None,
+        module: Optional[Module] = None,
+        variadic: bool = False,
+        initializer_list_func: Optional[Callable[[Dict[str, Any], type], bool]] = None,
+        export: bool = False,
+        doc: str = "",
+        group: str = "",
+        hidden: bool = False,
+        skip_replay: bool = False,
+        missing_grad: bool = False,
+        generic: bool = False,
+        native_func: Optional[str] = None,
+        defaults: Optional[Dict[str, Any]] = None,
+        custom_replay_func: Optional[Function] = None,
+        native_snippet: Optional[str] = None,
+        adj_native_snippet: Optional[str] = None,
+        replay_snippet: Optional[str] = None,
+        skip_forward_codegen: bool = False,
+        skip_reverse_codegen: bool = False,
+        custom_reverse_num_input_args: int = -1,
+        custom_reverse_mode: bool = False,
+        overloaded_annotations: Optional[Dict[str, type]] = None,
+        code_transformers: Optional[List[ast.NodeTransformer]] = None,
+        skip_adding_overload: bool = False,
+        require_original_output_arg: bool = False,
+        scope_locals: Optional[Dict[str, Any]] = None,
     ):
         if code_transformers is None:
             code_transformers = []
@@ -165,7 +178,7 @@ class Function:
         self.native_snippet = native_snippet
         self.adj_native_snippet = adj_native_snippet
         self.replay_snippet = replay_snippet
-        self.custom_grad_func = None
+        self.custom_grad_func: Optional[Function] = None
         self.require_original_output_arg = require_original_output_arg
         self.generic_parent = None  # generic function that was used to instantiate this overload
 
@@ -181,6 +194,7 @@ class Function:
         )
         self.missing_grad = missing_grad  # whether builtin is missing a corresponding adjoint
         self.generic = generic
+        self.mangled_name: Optional[str] = None
 
         # allow registering functions with a different name in Python and native code
         if native_func is None:
@@ -197,8 +211,8 @@ class Function:
             # user-defined function
 
             # generic and concrete overload lookups by type signature
-            self.user_templates = {}
-            self.user_overloads = {}
+            self.user_templates: Dict[str, Function] = {}
+            self.user_overloads: Dict[str, Function] = {}
 
             # user defined (Python) function
             self.adj = warp.codegen.Adjoint(
@@ -229,19 +243,17 @@ class Function:
             # builtin function
 
             # embedded linked list of all overloads
-            # the builtin_functions dictionary holds
-            # the list head for a given key (func name)
-            self.overloads = []
+            # the builtin_functions dictionary holds the list head for a given key (func name)
+            self.overloads: List[Function] = []
 
             # builtin (native) function, canonicalize argument types
-            for k, v in input_types.items():
-                self.input_types[k] = warp.types.type_to_warp(v)
+            if input_types is not None:
+                for k, v in input_types.items():
+                    self.input_types[k] = warp.types.type_to_warp(v)
 
             # cache mangled name
             if self.export and self.is_simple():
                 self.mangled_name = self.mangle()
-            else:
-                self.mangled_name = None
 
         if not skip_adding_overload:
             self.add_overload(self)
@@ -272,7 +284,7 @@ class Function:
             signature_params.append(param)
         self.signature = inspect.Signature(signature_params)
 
-        # scope for resolving overloads
+        # scope for resolving overloads, the locals() where the function is defined
         if scope_locals is None:
             scope_locals = inspect.currentframe().f_back.f_locals
 
@@ -334,10 +346,10 @@ class Function:
         # this function has no overloads, call it like a plain Python function
         return self.func(*args, **kwargs)
 
-    def is_builtin(self):
+    def is_builtin(self) -> bool:
         return self.func is None
 
-    def is_simple(self):
+    def is_simple(self) -> bool:
         if self.variadic:
             return False
 
@@ -351,9 +363,8 @@ class Function:
 
         return True
 
-    def mangle(self):
-        # builds a mangled name for the C-exported
-        # function, e.g.: builtin_normalize_vec3()
+    def mangle(self) -> str:
+        """Build a mangled name for the C-exported function, e.g.: `builtin_normalize_vec3()`."""
 
         name = "builtin_" + self.key
 
@@ -369,7 +380,7 @@ class Function:
 
         return "_".join([name, *types])
 
-    def add_overload(self, f):
+    def add_overload(self, f: Function) -> None:
         if self.is_builtin():
             # todo: note that it is an error to add two functions
             # with the exact same signature as this would cause compile
@@ -384,7 +395,7 @@ class Function:
         else:
             # get function signature based on the input types
             sig = warp.types.get_signature(
-                f.input_types.values(), func_name=f.key, arg_names=list(f.input_types.keys())
+                list(f.input_types.values()), func_name=f.key, arg_names=list(f.input_types.keys())
             )
 
             # check if generic
@@ -393,7 +404,7 @@ class Function:
             else:
                 self.user_overloads[sig] = f
 
-    def get_overload(self, arg_types, kwarg_types):
+    def get_overload(self, arg_types: List[type], kwarg_types: Mapping[str, type]) -> Optional[Function]:
         assert not self.is_builtin()
 
         for f in self.user_overloads.values():
@@ -446,7 +457,7 @@ class Function:
         return f"<Function {self.key}({inputs_str})>"
 
 
-def call_builtin(func: Function, *params) -> Tuple[bool, Any]:
+def call_builtin(func: Function, *params: Any) -> Tuple[bool, Any]:
     uses_non_warp_array_type = False
 
     init()
@@ -763,37 +774,51 @@ class Kernel:
 
 
 # decorator to register function, @func
-def func(f):
-    name = warp.codegen.make_full_qualified_name(f)
-
-    scope_locals = inspect.currentframe().f_back.f_locals
-
-    m = get_module(f.__module__)
-    doc = getattr(f, "__doc__", "") or ""
-    Function(
-        func=f,
-        key=name,
-        namespace="",
-        module=m,
-        value_func=None,
-        scope_locals=scope_locals,
-        doc=doc.strip(),
-    )  # value_type not known yet, will be inferred during Adjoint.build()
-
-    # use the top of the list of overloads for this key
-    g = m.functions[name]
-    # copy over the function attributes, including docstring
-    return functools.update_wrapper(g, f)
-
-
-def func_native(snippet, adj_snippet=None, replay_snippet=None):
+def func(f: Optional[Callable] = None, *, name: Optional[str] = None):
+    def wrapper(f, *args, **kwargs):
+        if name is None:
+            key = warp.codegen.make_full_qualified_name(f)
+        else:
+            key = name
+
+        scope_locals = inspect.currentframe().f_back.f_back.f_locals
+
+        m = get_module(f.__module__)
+        doc = getattr(f, "__doc__", "") or ""
+        Function(
+            func=f,
+            key=key,
+            namespace="",
+            module=m,
+            value_func=None,
+            scope_locals=scope_locals,
+            doc=doc.strip(),
+        )  # value_type not known yet, will be inferred during Adjoint.build()
+
+        # use the top of the list of overloads for this key
+        g = m.functions[key]
+        # copy over the function attributes, including docstring
+        return functools.update_wrapper(g, f)
+
+    if f is None:
+        # Arguments were passed to the decorator.
+        return wrapper
+
+    return wrapper(f)
+
+
+def func_native(snippet: str, adj_snippet: Optional[str] = None, replay_snippet: Optional[str] = None):
     """
     Decorator to register native code snippet, @func_native
     """
 
-    scope_locals = inspect.currentframe().f_back.f_locals
+    frame = inspect.currentframe()
+    if frame is None or frame.f_back is None:
+        scope_locals = {}
+    else:
+        scope_locals = frame.f_back.f_locals
 
-    def snippet_func(f):
+    def snippet_func(f: Callable) -> Callable:
         name = warp.codegen.make_full_qualified_name(f)
 
         m = get_module(f.__module__)
@@ -965,22 +990,71 @@ def func_replay(forward_fn):
     return wrapper
 
 
-# decorator to register kernel, @kernel, custom_name may be a string
-# that creates a kernel with a different name from the actual function
-def kernel(f=None, *, enable_backward=None):
+def kernel(
+    f: Optional[Callable] = None,
+    *,
+    enable_backward: Optional[bool] = None,
+    module: Optional[Union[Module, Literal["unique"]]] = None,
+):
+    """
+    Decorator to register a Warp kernel from a Python function.
+    The function must be defined with type annotations for all arguments.
+    The function must not return anything.
+
+    Example::
+
+        @wp.kernel
+        def my_kernel(a: wp.array(dtype=float), b: wp.array(dtype=float)):
+            tid = wp.tid()
+            b[tid] = a[tid] + 1.0
+
+
+        @wp.kernel(enable_backward=False)
+        def my_kernel_no_backward(a: wp.array(dtype=float, ndim=2), x: float):
+            # the backward pass will not be generated
+            i, j = wp.tid()
+            a[i, j] = x
+
+
+        @wp.kernel(module="unique")
+        def my_kernel_unique_module(a: wp.array(dtype=float), b: wp.array(dtype=float)):
+            # the kernel will be registered in new unique module created just for this
+            # kernel and its dependent functions and structs
+            tid = wp.tid()
+            b[tid] = a[tid] + 1.0
+
+    Args:
+        f: The function to be registered as a kernel.
+        enable_backward: If False, the backward pass will not be generated.
+        module: The :class:`warp.context.Module` to which the kernel belongs. Alternatively, if a string `"unique"` is provided, the kernel is assigned to a new module named after the kernel name and hash. If None, the module is inferred from the function's module.
+
+    Returns:
+        The registered kernel.
+    """
+
     def wrapper(f, *args, **kwargs):
         options = {}
 
         if enable_backward is not None:
             options["enable_backward"] = enable_backward
 
-        m = get_module(f.__module__)
+        if module is None:
+            m = get_module(f.__module__)
+        elif module == "unique":
+            m = Module(f.__name__, None)
+        else:
+            m = module
         k = Kernel(
             func=f,
             key=warp.codegen.make_full_qualified_name(f),
             module=m,
             options=options,
         )
+        if module == "unique":
+            # add the hash to the module name
+            hasher = warp.context.ModuleHasher(m)
+            k.module.name = f"{k.key}_{hasher.module_hash.hex()[:8]}"
+
         k = functools.update_wrapper(k, f)
         return k
 
@@ -992,7 +1066,7 @@ def kernel(f=None, *, enable_backward=None):
 
 
 # decorator to register struct, @struct
-def struct(c):
+def struct(c: type):
     m = get_module(c.__module__)
     s = warp.codegen.Struct(cls=c, key=warp.codegen.make_full_qualified_name(c), module=m)
     s = functools.update_wrapper(s, c)
@@ -1105,47 +1179,47 @@ scalar_types.update({x: x._wp_scalar_type_ for x in warp.types.vector_types})
 
 
 def add_builtin(
-    key,
-    input_types=None,
-    constraint=None,
-    value_type=None,
-    value_func=None,
-    export_func=None,
-    dispatch_func=None,
-    lto_dispatch_func=None,
-    doc="",
-    namespace="wp::",
-    variadic=False,
+    key: str,
+    input_types: Optional[Dict[str, Union[type, TypeVar]]] = None,
+    constraint: Optional[Callable[[Mapping[str, type]], bool]] = None,
+    value_type: Optional[type] = None,
+    value_func: Optional[Callable] = None,
+    export_func: Optional[Callable] = None,
+    dispatch_func: Optional[Callable] = None,
+    lto_dispatch_func: Optional[Callable] = None,
+    doc: str = "",
+    namespace: str = "wp::",
+    variadic: bool = False,
     initializer_list_func=None,
-    export=True,
-    group="Other",
-    hidden=False,
-    skip_replay=False,
-    missing_grad=False,
-    native_func=None,
-    defaults=None,
-    require_original_output_arg=False,
+    export: bool = True,
+    group: str = "Other",
+    hidden: bool = False,
+    skip_replay: bool = False,
+    missing_grad: bool = False,
+    native_func: Optional[str] = None,
+    defaults: Optional[Dict[str, Any]] = None,
+    require_original_output_arg: bool = False,
 ):
     """Main entry point to register a new built-in function.
 
     Args:
-        key (str): Function name. Multiple overloaded functions can be registered
+        key: Function name. Multiple overloaded functions can be registered
             under the same name as long as their signature differ.
-        input_types (Mapping[str, Any]): Signature of the user-facing function.
+        input_types: Signature of the user-facing function.
             Variadic arguments are supported by prefixing the parameter names
             with asterisks as in `*args` and `**kwargs`. Generic arguments are
             supported with types such as `Any`, `Float`, `Scalar`, etc.
-        constraint (Callable): For functions that define generic arguments and
+        constraint: For functions that define generic arguments and
             are to be exported, this callback is used to specify whether some
             combination of inferred arguments are valid or not.
-        value_type (Any): Type returned by the function.
-        value_func (Callable): Callback used to specify the return type when
+        value_type: Type returned by the function.
+        value_func: Callback used to specify the return type when
             `value_type` isn't enough.
-        export_func (Callable): Callback used during the context stage to specify
+        export_func: Callback used during the context stage to specify
             the signature of the underlying C++ function, not accounting for
             the template parameters.
             If not provided, `input_types` is used.
-        dispatch_func (Callable): Callback used during the codegen stage to specify
+        dispatch_func: Callback used during the codegen stage to specify
             the runtime and template arguments to be passed to the underlying C++
             function. In other words, this allows defining a mapping between
             the signatures of the user-facing and the C++ functions, and even to
@@ -1153,27 +1227,26 @@ def add_builtin(
             The arguments returned must be of type `codegen.Var`.
             If not provided, all arguments passed by the users when calling
             the built-in are passed as-is as runtime arguments to the C++ function.
-        lto_dispatch_func (Callable): Same as dispatch_func, but takes an 'option' dict
+        lto_dispatch_func: Same as dispatch_func, but takes an 'option' dict
             as extra argument (indicating tile_size and target architecture) and returns
             an LTO-IR buffer as extra return value
-        doc (str): Used to generate the Python's docstring and the HTML documentation.
+        doc: Used to generate the Python's docstring and the HTML documentation.
         namespace: Namespace for the underlying C++ function.
-        variadic (bool): Whether the function declares variadic arguments.
-        initializer_list_func (bool): Whether to use the initializer list syntax
-            when passing the arguments to the underlying C++ function.
-        export (bool): Whether the function is to be exposed to the Python
+        variadic: Whether the function declares variadic arguments.
+        initializer_list_func: Callback to determine whether to use the
+            initializer list syntax when passing the arguments to the underlying
+            C++ function.
+        export: Whether the function is to be exposed to the Python
             interpreter so that it becomes available from within the `warp`
             module.
-        group (str): Classification used for the documentation.
-        hidden (bool): Whether to add that function into the documentation.
-        skip_replay (bool): Whether operation will be performed during
+        group: Classification used for the documentation.
+        hidden: Whether to add that function into the documentation.
+        skip_replay: Whether operation will be performed during
             the forward replay in the backward pass.
-        missing_grad (bool): Whether the function is missing a corresponding
-            adjoint.
-        native_func (str): Name of the underlying C++ function.
-        defaults (Mapping[str, Any]): Default values for the parameters defined
-            in `input_types`.
-        require_original_output_arg (bool): Used during the codegen stage to
+        missing_grad: Whether the function is missing a corresponding adjoint.
+        native_func: Name of the underlying C++ function.
+        defaults: Default values for the parameters defined in `input_types`.
+        require_original_output_arg: Used during the codegen stage to
             specify whether an adjoint parameter corresponding to the return
             value should be included in the signature of the backward function.
     """
@@ -1355,19 +1428,14 @@ def add_builtin(
 def register_api_function(
     function: Function,
     group: str = "Other",
-    hidden=False,
+    hidden: bool = False,
 ):
     """Main entry point to register a Warp Python function to be part of the Warp API and appear in the documentation.
 
     Args:
-        function (Function): Warp function to be registered.
-        group (str): Classification used for the documentation.
-        input_types (Mapping[str, Any]): Signature of the user-facing function.
-            Variadic arguments are supported by prefixing the parameter names
-            with asterisks as in `*args` and `**kwargs`. Generic arguments are
-            supported with types such as `Any`, `Float`, `Scalar`, etc.
-        value_type (Any): Type returned by the function.
-        hidden (bool): Whether to add that function into the documentation.
+        function: Warp function to be registered.
+        group: Classification used for the documentation.
+        hidden: Whether to add that function into the documentation.
     """
     function.group = group
     function.hidden = hidden
@@ -1375,10 +1443,10 @@ def register_api_function(
 
 
 # global dictionary of modules
-user_modules = {}
+user_modules: Dict[str, Module] = {}
 
 
-def get_module(name):
+def get_module(name: str) -> Module:
     # some modules might be manually imported using `importlib` without being
     # registered into `sys.modules`
     parent = sys.modules.get(name, None)
@@ -1460,13 +1528,16 @@ class ModuleHasher:
         if warp.config.verify_fp:
             ch.update(bytes("verify_fp", "utf-8"))
 
+        # line directives, e.g. for Nsight Compute
+        ch.update(bytes(ctypes.c_int(warp.config.line_directives)))
+
         # build config
         ch.update(bytes(warp.config.mode, "utf-8"))
 
         # save the module hash
         self.module_hash = ch.digest()
 
-    def hash_kernel(self, kernel):
+    def hash_kernel(self, kernel: Kernel) -> bytes:
         # NOTE: We only hash non-generic kernels, so we don't traverse kernel overloads here.
 
         ch = hashlib.sha256()
@@ -1480,7 +1551,7 @@ class ModuleHasher:
 
         return h
 
-    def hash_function(self, func):
+    def hash_function(self, func: Function) -> bytes:
         # NOTE: This method hashes all possible overloads that a function call could resolve to.
         # The exact overload will be resolved at build time, when the argument types are known.
 
@@ -1495,7 +1566,7 @@ class ModuleHasher:
         ch.update(bytes(func.key, "utf-8"))
 
         # include all concrete and generic overloads
-        overloads = {**func.user_overloads, **func.user_templates}
+        overloads: Dict[str, Function] = {**func.user_overloads, **func.user_templates}
         for sig in sorted(overloads.keys()):
             ovl = overloads[sig]
 
@@ -1526,7 +1597,7 @@ class ModuleHasher:
 
         return h
 
-    def hash_adjoint(self, adj):
+    def hash_adjoint(self, adj: warp.codegen.Adjoint) -> bytes:
         # NOTE: We don't cache adjoint hashes, because adjoints are always unique.
         # Even instances of generic kernels and functions have unique adjoints with
         # different argument types.
@@ -1575,7 +1646,7 @@ class ModuleHasher:
 
         return ch.digest()
 
-    def get_constant_bytes(self, value):
+    def get_constant_bytes(self, value) -> bytes:
         if isinstance(value, int):
             # this also handles builtins.bool
             return bytes(ctypes.c_int(value))
@@ -1593,7 +1664,7 @@ class ModuleHasher:
         else:
             raise TypeError(f"Invalid constant type: {type(value)}")
 
-    def get_module_hash(self):
+    def get_module_hash(self) -> bytes:
         return self.module_hash
 
     def get_unique_kernels(self):
@@ -1610,6 +1681,7 @@ class ModuleBuilder:
         self.fatbins = {}  # map from <some identifier> to fatbins, to add at link time
         self.ltoirs = {}  # map from lto symbol to lto binary
         self.ltoirs_decl = {}  # map from lto symbol to lto forward declaration
+        self.shared_memory_bytes = {}  # map from lto symbol to shared memory requirements
 
         if hasher is None:
             hasher = ModuleHasher(module)
@@ -1726,9 +1798,9 @@ class ModuleBuilder:
 
         # add headers
         if device == "cpu":
-            source = warp.codegen.cpu_module_header.format(tile_size=self.options["block_dim"]) + source
+            source = warp.codegen.cpu_module_header.format(block_dim=self.options["block_dim"]) + source
         else:
-            source = warp.codegen.cuda_module_header.format(tile_size=self.options["block_dim"]) + source
+            source = warp.codegen.cuda_module_header.format(block_dim=self.options["block_dim"]) + source
 
         return source
 
@@ -1765,7 +1837,7 @@ class ModuleExec:
                 runtime.llvm.unload_obj(self.handle.encode("utf-8"))
 
     # lookup and cache kernel entry points
-    def get_kernel_hooks(self, kernel):
+    def get_kernel_hooks(self, kernel) -> KernelHooks:
         # Use kernel.adj as a unique key for cache lookups instead of the kernel itself.
         # This avoids holding a reference to the kernel and is faster than using
         # a WeakKeyDictionary with kernels as keys.
@@ -1838,7 +1910,7 @@ class ModuleExec:
 # creates a hash of the function to use for checking
 # build cache
 class Module:
-    def __init__(self, name, loader):
+    def __init__(self, name: Optional[str], loader=None):
         self.name = name if name is not None else "None"
 
         self.loader = loader
@@ -1878,7 +1950,7 @@ class Module:
             "enable_backward": warp.config.enable_backward,
             "fast_math": False,
             "fuse_fp": True,
-            "lineinfo": False,
+            "lineinfo": warp.config.lineinfo,
             "cuda_output": None,  # supported values: "ptx", "cubin", or None (automatic)
             "mode": warp.config.mode,
             "block_dim": 256,
@@ -2081,7 +2153,11 @@ class Module:
                     use_ptx = True
 
                 if use_ptx:
-                    output_arch = min(device.arch, warp.config.ptx_target_arch)
+                    # use the default PTX arch if the device supports it
+                    if warp.config.ptx_target_arch is not None:
+                        output_arch = min(device.arch, warp.config.ptx_target_arch)
+                    else:
+                        output_arch = min(device.arch, runtime.default_ptx_arch)
                     output_name = f"{module_name_short}.sm{output_arch}.ptx"
                 else:
                     output_arch = device.arch
@@ -2194,34 +2270,8 @@ class Module:
                 # -----------------------------------------------------------
                 # update cache
 
-                def safe_rename(src, dst, attempts=5, delay=0.1):
-                    for i in range(attempts):
-                        try:
-                            os.rename(src, dst)
-                            return
-                        except FileExistsError:
-                            return
-                        except OSError as e:
-                            if e.errno == errno.ENOTEMPTY:
-                                # if directory exists we assume another process
-                                # got there first, in which case we will copy
-                                # our output to the directory manually in second step
-                                return
-                            else:
-                                # otherwise assume directory creation failed e.g.: access denied
-                                # on Windows we see occasional failures to rename directories due to
-                                # some process holding a lock on a file to be moved to workaround
-                                # this we make multiple attempts to rename with some delay
-                                if i < attempts - 1:
-                                    time.sleep(delay)
-                                else:
-                                    print(
-                                        f"Could not update Warp cache with module binaries, trying to rename {build_dir} to {module_dir}, error {e}"
-                                    )
-                                    raise e
-
                 # try to move process outputs to cache
-                safe_rename(build_dir, module_dir)
+                warp.build.safe_rename(build_dir, module_dir)
 
                 if os.path.exists(module_dir):
                     if not os.path.exists(binary_path):
@@ -2294,7 +2344,7 @@ class Module:
         self.failed_builds = set()
 
     # lookup kernel entry points based on name, called after compilation / module load
-    def get_kernel_hooks(self, kernel, device):
+    def get_kernel_hooks(self, kernel, device: Device) -> KernelHooks:
         module_exec = self.execs.get((device.context, self.options["block_dim"]))
         if module_exec is not None:
             return module_exec.get_kernel_hooks(kernel)
@@ -2449,6 +2499,7 @@ class Event:
             raise RuntimeError(f"Device {device} is not a CUDA device")
 
         self.device = device
+        self.enable_timing = enable_timing
 
         if cuda_event is not None:
             self.cuda_event = cuda_event
@@ -2498,6 +2549,17 @@ class Event:
         else:
             raise RuntimeError(f"Device {self.device} does not support IPC.")
 
+    @property
+    def is_complete(self) -> bool:
+        """A boolean indicating whether all work on the stream when the event was recorded has completed.
+
+        This property may not be accessed during a graph capture on any stream.
+        """
+
+        result_code = runtime.core.cuda_event_query(self.cuda_event)
+
+        return result_code == 0
+
     def __del__(self):
         if not self.owner:
             return
@@ -2512,7 +2574,7 @@ class Stream:
         instance.owner = False
         return instance
 
-    def __init__(self, device: Optional[Union["Device", str]] = None, priority: int = 0, **kwargs):
+    def __init__(self, device: Union["Device", str, None] = None, priority: int = 0, **kwargs):
         """Initialize the stream on a device with an optional specified priority.
 
         Args:
@@ -2528,7 +2590,7 @@ class Stream:
         Raises:
             RuntimeError: If function is called before Warp has completed
               initialization with a ``device`` that is not an instance of
-              :class:`Device``.
+              :class:`Device <warp.context.Device>`.
             RuntimeError: ``device`` is not a CUDA Device.
             RuntimeError: The stream could not be created on the device.
             TypeError: The requested stream priority is not an integer.
@@ -2596,7 +2658,7 @@ class Stream:
                 f"Event from device {event.device} cannot be recorded on stream from device {self.device}"
             )
 
-        runtime.core.cuda_event_record(event.cuda_event, self.cuda_stream)
+        runtime.core.cuda_event_record(event.cuda_event, self.cuda_stream, event.enable_timing)
 
         return event
 
@@ -2630,6 +2692,17 @@ class Stream:
 
         runtime.core.cuda_stream_wait_stream(self.cuda_stream, other_stream.cuda_stream, event.cuda_event)
 
+    @property
+    def is_complete(self) -> bool:
+        """A boolean indicating whether all work on the stream has completed.
+
+        This property may not be accessed during a graph capture on any stream.
+        """
+
+        result_code = runtime.core.cuda_stream_query(self.cuda_stream)
+
+        return result_code == 0
+
     @property
     def is_capturing(self) -> bool:
         """A boolean indicating whether a graph capture is currently ongoing on this stream."""
@@ -2952,18 +3025,14 @@ Devicelike = Union[Device, str, None]
 
 
 class Graph:
-    def __new__(cls, *args, **kwargs):
-        instance = super(Graph, cls).__new__(cls)
-        instance.graph_exec = None
-        return instance
-
     def __init__(self, device: Device, capture_id: int):
         self.device = device
         self.capture_id = capture_id
-        self.module_execs = set()
+        self.module_execs: Set[ModuleExec] = set()
+        self.graph_exec: Optional[ctypes.c_void_p] = None
 
     def __del__(self):
-        if not self.graph_exec:
+        if not hasattr(self, "graph_exec") or not hasattr(self, "device") or not self.graph_exec:
             return
 
         # use CUDA context guard to avoid side effects during garbage collection
@@ -3205,6 +3274,43 @@ class Runtime:
             self.core.radix_sort_pairs_float_host.argtypes = [ctypes.c_uint64, ctypes.c_uint64, ctypes.c_int]
             self.core.radix_sort_pairs_float_device.argtypes = [ctypes.c_uint64, ctypes.c_uint64, ctypes.c_int]
 
+            self.core.radix_sort_pairs_int64_host.argtypes = [ctypes.c_uint64, ctypes.c_uint64, ctypes.c_int]
+            self.core.radix_sort_pairs_int64_device.argtypes = [ctypes.c_uint64, ctypes.c_uint64, ctypes.c_int]
+
+            self.core.segmented_sort_pairs_int_host.argtypes = [
+                ctypes.c_uint64,
+                ctypes.c_uint64,
+                ctypes.c_int,
+                ctypes.c_uint64,
+                ctypes.c_uint64,
+                ctypes.c_int,
+            ]
+            self.core.segmented_sort_pairs_int_device.argtypes = [
+                ctypes.c_uint64,
+                ctypes.c_uint64,
+                ctypes.c_int,
+                ctypes.c_uint64,
+                ctypes.c_uint64,
+                ctypes.c_int,
+            ]
+
+            self.core.segmented_sort_pairs_float_host.argtypes = [
+                ctypes.c_uint64,
+                ctypes.c_uint64,
+                ctypes.c_int,
+                ctypes.c_uint64,
+                ctypes.c_uint64,
+                ctypes.c_int,
+            ]
+            self.core.segmented_sort_pairs_float_device.argtypes = [
+                ctypes.c_uint64,
+                ctypes.c_uint64,
+                ctypes.c_int,
+                ctypes.c_uint64,
+                ctypes.c_uint64,
+                ctypes.c_int,
+            ]
+
             self.core.runlength_encode_int_host.argtypes = [
                 ctypes.c_uint64,
                 ctypes.c_uint64,
@@ -3285,26 +3391,6 @@ class Runtime:
             self.core.hash_grid_update_device.argtypes = [ctypes.c_uint64, ctypes.c_float, ctypes.c_void_p]
             self.core.hash_grid_reserve_device.argtypes = [ctypes.c_uint64, ctypes.c_int]
 
-            self.core.cutlass_gemm.argtypes = [
-                ctypes.c_void_p,
-                ctypes.c_int,
-                ctypes.c_int,
-                ctypes.c_int,
-                ctypes.c_int,
-                ctypes.c_char_p,
-                ctypes.c_void_p,
-                ctypes.c_void_p,
-                ctypes.c_void_p,
-                ctypes.c_void_p,
-                ctypes.c_float,
-                ctypes.c_float,
-                ctypes.c_bool,
-                ctypes.c_bool,
-                ctypes.c_bool,
-                ctypes.c_int,
-            ]
-            self.core.cutlass_gemm.restype = ctypes.c_bool
-
             self.core.volume_create_host.argtypes = [ctypes.c_void_p, ctypes.c_uint64, ctypes.c_bool, ctypes.c_bool]
             self.core.volume_create_host.restype = ctypes.c_uint64
             self.core.volume_get_tiles_host.argtypes = [
@@ -3335,36 +3421,18 @@ class Runtime:
             ]
             self.core.volume_destroy_device.argtypes = [ctypes.c_uint64]
 
-            self.core.volume_f_from_tiles_device.argtypes = [
+            self.core.volume_from_tiles_device.argtypes = [
                 ctypes.c_void_p,
                 ctypes.c_void_p,
                 ctypes.c_int,
                 ctypes.c_float * 9,
                 ctypes.c_float * 3,
                 ctypes.c_bool,
-                ctypes.c_float,
-            ]
-            self.core.volume_f_from_tiles_device.restype = ctypes.c_uint64
-            self.core.volume_v_from_tiles_device.argtypes = [
-                ctypes.c_void_p,
                 ctypes.c_void_p,
-                ctypes.c_int,
-                ctypes.c_float * 9,
-                ctypes.c_float * 3,
-                ctypes.c_bool,
-                ctypes.c_float * 3,
-            ]
-            self.core.volume_v_from_tiles_device.restype = ctypes.c_uint64
-            self.core.volume_i_from_tiles_device.argtypes = [
-                ctypes.c_void_p,
-                ctypes.c_void_p,
-                ctypes.c_int,
-                ctypes.c_float * 9,
-                ctypes.c_float * 3,
-                ctypes.c_bool,
-                ctypes.c_int,
+                ctypes.c_uint32,
+                ctypes.c_char_p,
             ]
-            self.core.volume_i_from_tiles_device.restype = ctypes.c_uint64
+            self.core.volume_from_tiles_device.restype = ctypes.c_uint64
             self.core.volume_index_from_tiles_device.argtypes = [
                 ctypes.c_void_p,
                 ctypes.c_void_p,
@@ -3433,6 +3501,7 @@ class Runtime:
                 ctypes.POINTER(ctypes.c_int),  # tpl_cols
                 ctypes.c_void_p,  # tpl_values
                 ctypes.c_bool,  # prune_numerical_zeros
+                ctypes.c_bool,  # masked
                 ctypes.POINTER(ctypes.c_int),  # bsr_offsets
                 ctypes.POINTER(ctypes.c_int),  # bsr_columns
                 ctypes.c_void_p,  # bsr_values
@@ -3467,8 +3536,6 @@ class Runtime:
             self.core.is_cuda_enabled.restype = ctypes.c_int
             self.core.is_cuda_compatibility_enabled.argtypes = None
             self.core.is_cuda_compatibility_enabled.restype = ctypes.c_int
-            self.core.is_cutlass_enabled.argtypes = None
-            self.core.is_cutlass_enabled.restype = ctypes.c_int
             self.core.is_mathdx_enabled.argtypes = None
             self.core.is_mathdx_enabled.restype = ctypes.c_int
 
@@ -3502,6 +3569,10 @@ class Runtime:
             self.core.cuda_device_set_mempool_release_threshold.restype = ctypes.c_int
             self.core.cuda_device_get_mempool_release_threshold.argtypes = [ctypes.c_int]
             self.core.cuda_device_get_mempool_release_threshold.restype = ctypes.c_uint64
+            self.core.cuda_device_get_mempool_used_mem_current.argtypes = [ctypes.c_int]
+            self.core.cuda_device_get_mempool_used_mem_current.restype = ctypes.c_uint64
+            self.core.cuda_device_get_mempool_used_mem_high.argtypes = [ctypes.c_int]
+            self.core.cuda_device_get_mempool_used_mem_high.restype = ctypes.c_uint64
             self.core.cuda_device_get_memory_info.argtypes = [ctypes.c_int, ctypes.c_void_p, ctypes.c_void_p]
             self.core.cuda_device_get_memory_info.restype = None
             self.core.cuda_device_get_uuid.argtypes = [ctypes.c_int, ctypes.c_char * 16]
@@ -3571,6 +3642,8 @@ class Runtime:
             self.core.cuda_stream_create.restype = ctypes.c_void_p
             self.core.cuda_stream_destroy.argtypes = [ctypes.c_void_p, ctypes.c_void_p]
             self.core.cuda_stream_destroy.restype = None
+            self.core.cuda_stream_query.argtypes = [ctypes.c_void_p]
+            self.core.cuda_stream_query.restype = ctypes.c_int
             self.core.cuda_stream_register.argtypes = [ctypes.c_void_p, ctypes.c_void_p]
             self.core.cuda_stream_register.restype = None
             self.core.cuda_stream_unregister.argtypes = [ctypes.c_void_p, ctypes.c_void_p]
@@ -3592,7 +3665,9 @@ class Runtime:
             self.core.cuda_event_create.restype = ctypes.c_void_p
             self.core.cuda_event_destroy.argtypes = [ctypes.c_void_p]
             self.core.cuda_event_destroy.restype = None
-            self.core.cuda_event_record.argtypes = [ctypes.c_void_p, ctypes.c_void_p]
+            self.core.cuda_event_query.argtypes = [ctypes.c_void_p]
+            self.core.cuda_event_query.restype = ctypes.c_int
+            self.core.cuda_event_record.argtypes = [ctypes.c_void_p, ctypes.c_void_p, ctypes.c_bool]
             self.core.cuda_event_record.restype = None
             self.core.cuda_event_synchronize.argtypes = [ctypes.c_void_p]
             self.core.cuda_event_synchronize.restype = None
@@ -3841,9 +3916,20 @@ class Runtime:
                 cuda_device_count = len(self.cuda_devices)
             else:
                 self.set_default_device("cuda:0")
+
+            # the minimum PTX architecture that supports all of Warp's features
+            self.default_ptx_arch = 75
+
+            # Update the default PTX architecture based on devices present in the system.
+            # Use the lowest architecture among devices that meet the minimum architecture requirement.
+            # Devices below the required minimum will use the highest architecture they support.
+            eligible_archs = [d.arch for d in self.cuda_devices if d.arch >= self.default_ptx_arch]
+            if eligible_archs:
+                self.default_ptx_arch = min(eligible_archs)
         else:
             # CUDA not available
             self.set_default_device("cpu")
+            self.default_ptx_arch = None
 
         # initialize kernel cache
         warp.build.init_kernel_cache(warp.config.kernel_cache_dir)
@@ -3856,6 +3942,11 @@ class Runtime:
             greeting = []
 
             greeting.append(f"Warp {warp.config.version} initialized:")
+
+            # Add git commit hash to greeting if available
+            if warp.config._git_commit_hash is not None:
+                greeting.append(f"   Git commit: {warp.config._git_commit_hash}")
+
             if cuda_device_count > 0:
                 # print CUDA version info
                 greeting.append(
@@ -4208,7 +4299,7 @@ def set_device(ident: Devicelike) -> None:
     device.make_current()
 
 
-def map_cuda_device(alias: str, context: ctypes.c_void_p = None) -> Device:
+def map_cuda_device(alias: str, context: Optional[ctypes.c_void_p] = None) -> Device:
     """Assign a device alias to a CUDA context.
 
     This function can be used to create a wp.Device for an external CUDA context.
@@ -4236,7 +4327,13 @@ def unmap_cuda_device(alias: str) -> None:
 
 
 def is_mempool_supported(device: Devicelike) -> bool:
-    """Check if CUDA memory pool allocators are available on the device."""
+    """Check if CUDA memory pool allocators are available on the device.
+
+    Parameters:
+        device: The :class:`Device <warp.context.Device>` or device identifier
+          for which the query is to be performed.
+          If ``None``, the default device will be used.
+    """
 
     init()
 
@@ -4246,7 +4343,13 @@ def is_mempool_supported(device: Devicelike) -> bool:
 
 
 def is_mempool_enabled(device: Devicelike) -> bool:
-    """Check if CUDA memory pool allocators are enabled on the device."""
+    """Check if CUDA memory pool allocators are enabled on the device.
+
+    Parameters:
+        device: The :class:`Device <warp.context.Device>` or device identifier
+          for which the query is to be performed.
+          If ``None``, the default device will be used.
+    """
 
     init()
 
@@ -4266,6 +4369,11 @@ def set_mempool_enabled(device: Devicelike, enable: bool) -> None:
     to Warp.  The preferred solution is to enable memory pool access using :func:`set_mempool_access_enabled`.
     If peer access is not supported, then the default CUDA allocators must be used to pre-allocate the memory
     prior to graph capture.
+
+    Parameters:
+        device: The :class:`Device <warp.context.Device>` or device identifier
+          for which the operation is to be performed.
+          If ``None``, the default device will be used.
     """
 
     init()
@@ -4296,6 +4404,18 @@ def set_mempool_release_threshold(device: Devicelike, threshold: Union[int, floa
     Values between 0 and 1 are interpreted as fractions of available memory.  For example, 0.5 means
     half of the device's physical memory.  Greater values are interpreted as an absolute number of bytes.
     For example, 1024**3 means one GiB of memory.
+
+    Parameters:
+        device: The :class:`Device <warp.context.Device>` or device identifier
+          for which the operation is to be performed.
+          If ``None``, the default device will be used.
+        threshold: An integer representing a number of bytes, or a ``float`` between 0 and 1,
+          specifying the desired release threshold.
+
+    Raises:
+        ValueError: If ``device`` is not a CUDA device.
+        RuntimeError: If ``device`` is a CUDA device, but does not support memory pools.
+        RuntimeError: Failed to set the memory pool release threshold.
     """
 
     init()
@@ -4317,8 +4437,21 @@ def set_mempool_release_threshold(device: Devicelike, threshold: Union[int, floa
         raise RuntimeError(f"Failed to set memory pool release threshold for device {device}")
 
 
-def get_mempool_release_threshold(device: Devicelike) -> int:
-    """Get the CUDA memory pool release threshold on the device in bytes."""
+def get_mempool_release_threshold(device: Devicelike = None) -> int:
+    """Get the CUDA memory pool release threshold on the device.
+
+    Parameters:
+        device: The :class:`Device <warp.context.Device>` or device identifier
+          for which the query is to be performed.
+          If ``None``, the default device will be used.
+
+    Returns:
+        The memory pool release threshold in bytes.
+
+    Raises:
+        ValueError: If ``device`` is not a CUDA device.
+        RuntimeError: If ``device`` is a CUDA device, but does not support memory pools.
+    """
 
     init()
 
@@ -4333,6 +4466,64 @@ def get_mempool_release_threshold(device: Devicelike) -> int:
     return runtime.core.cuda_device_get_mempool_release_threshold(device.ordinal)
 
 
+def get_mempool_used_mem_current(device: Devicelike = None) -> int:
+    """Get the amount of memory from the device's memory pool that is currently in use by the application.
+
+    Parameters:
+        device: The :class:`Device <warp.context.Device>` or device identifier
+          for which the query is to be performed.
+          If ``None``, the default device will be used.
+
+    Returns:
+        The amount of memory used in bytes.
+
+    Raises:
+        ValueError: If ``device`` is not a CUDA device.
+        RuntimeError: If ``device`` is a CUDA device, but does not support memory pools.
+    """
+
+    init()
+
+    device = runtime.get_device(device)
+
+    if not device.is_cuda:
+        raise ValueError("Memory pools are only supported on CUDA devices")
+
+    if not device.is_mempool_supported:
+        raise RuntimeError(f"Device {device} does not support memory pools")
+
+    return runtime.core.cuda_device_get_mempool_used_mem_current(device.ordinal)
+
+
+def get_mempool_used_mem_high(device: Devicelike = None) -> int:
+    """Get the application's memory usage high-water mark from the device's CUDA memory pool.
+
+    Parameters:
+        device: The :class:`Device <warp.context.Device>` or device identifier
+          for which the query is to be performed.
+          If ``None``, the default device will be used.
+
+    Returns:
+        The high-water mark of memory used from the memory pool in bytes.
+
+    Raises:
+        ValueError: If ``device`` is not a CUDA device.
+        RuntimeError: If ``device`` is a CUDA device, but does not support memory pools.
+    """
+
+    init()
+
+    device = runtime.get_device(device)
+
+    if not device.is_cuda:
+        raise ValueError("Memory pools are only supported on CUDA devices")
+
+    if not device.is_mempool_supported:
+        raise RuntimeError(f"Device {device} does not support memory pools")
+
+    return runtime.core.cuda_device_get_mempool_used_mem_high(device.ordinal)
+
+
 def is_peer_access_supported(target_device: Devicelike, peer_device: Devicelike) -> bool:
     """Check if `peer_device` can directly access the memory of `target_device` on this system.
 
@@ -4535,7 +4726,7 @@ def wait_event(event: Event):
     get_stream().wait_event(event)
 
 
-def get_event_elapsed_time(start_event: Event, end_event: Event, synchronize: Optional[bool] = True):
+def get_event_elapsed_time(start_event: Event, end_event: Event, synchronize: bool = True):
     """Get the elapsed time between two recorded events.
 
     Both events must have been previously recorded with
@@ -4560,7 +4751,7 @@ def get_event_elapsed_time(start_event: Event, end_event: Event, synchronize: Op
     return runtime.core.cuda_event_elapsed_time(start_event.cuda_event, end_event.cuda_event)
 
 
-def wait_stream(other_stream: Stream, event: Event = None):
+def wait_stream(other_stream: Stream, event: Optional[Event] = None):
     """Convenience function for calling :meth:`Stream.wait_stream` on the current stream.
 
     Args:
@@ -4727,7 +4918,7 @@ class RegisteredGLBuffer:
 
 
 def zeros(
-    shape: Tuple = None,
+    shape: Union[int, Tuple[int, ...], List[int], None] = None,
     dtype=float,
     device: Devicelike = None,
     requires_grad: bool = False,
@@ -4755,7 +4946,7 @@ def zeros(
 
 
 def zeros_like(
-    src: warp.array, device: Devicelike = None, requires_grad: bool = None, pinned: bool = None
+    src: Array, device: Devicelike = None, requires_grad: Optional[bool] = None, pinned: Optional[bool] = None
 ) -> warp.array:
     """Return a zero-initialized array with the same type and dimension of another array
 
@@ -4777,7 +4968,7 @@ def zeros_like(
 
 
 def ones(
-    shape: Tuple = None,
+    shape: Union[int, Tuple[int, ...], List[int], None] = None,
     dtype=float,
     device: Devicelike = None,
     requires_grad: bool = False,
@@ -4801,7 +4992,7 @@ def ones(
 
 
 def ones_like(
-    src: warp.array, device: Devicelike = None, requires_grad: bool = None, pinned: bool = None
+    src: Array, device: Devicelike = None, requires_grad: Optional[bool] = None, pinned: Optional[bool] = None
 ) -> warp.array:
     """Return a one-initialized array with the same type and dimension of another array
 
@@ -4819,7 +5010,7 @@ def ones_like(
 
 
 def full(
-    shape: Tuple = None,
+    shape: Union[int, Tuple[int, ...], List[int], None] = None,
     value=0,
     dtype=Any,
     device: Devicelike = None,
@@ -4885,7 +5076,11 @@ def full(
 
 
 def full_like(
-    src: warp.array, value: Any, device: Devicelike = None, requires_grad: bool = None, pinned: bool = None
+    src: Array,
+    value: Any,
+    device: Devicelike = None,
+    requires_grad: Optional[bool] = None,
+    pinned: Optional[bool] = None,
 ) -> warp.array:
     """Return an array with all elements initialized to the given value with the same type and dimension of another array
 
@@ -4907,7 +5102,9 @@ def full_like(
     return arr
 
 
-def clone(src: warp.array, device: Devicelike = None, requires_grad: bool = None, pinned: bool = None) -> warp.array:
+def clone(
+    src: warp.array, device: Devicelike = None, requires_grad: Optional[bool] = None, pinned: Optional[bool] = None
+) -> warp.array:
     """Clone an existing array, allocates a copy of the src memory
 
     Args:
@@ -4928,7 +5125,7 @@ def clone(src: warp.array, device: Devicelike = None, requires_grad: bool = None
 
 
 def empty(
-    shape: Tuple = None,
+    shape: Union[int, Tuple[int, ...], List[int], None] = None,
     dtype=float,
     device: Devicelike = None,
     requires_grad: bool = False,
@@ -4961,7 +5158,7 @@ def empty(
 
 
 def empty_like(
-    src: warp.array, device: Devicelike = None, requires_grad: bool = None, pinned: bool = None
+    src: Array, device: Devicelike = None, requires_grad: Optional[bool] = None, pinned: Optional[bool] = None
 ) -> warp.array:
     """Return an uninitialized array with the same type and dimension of another array
 
@@ -5193,8 +5390,6 @@ def pack_arg(kernel, arg_type, arg_name, value, device, adjoint=False):
             ) from e
 
 
-# represents all data required for a kernel launch
-# so that launches can be replayed quickly, use `wp.launch(..., record_cmd=True)`
 class Launch:
     """Represents all data required for a kernel launch so that launches can be replayed quickly.
 
@@ -5465,7 +5660,7 @@ def launch(
         max_blocks: The maximum number of CUDA thread blocks to use.
           Only has an effect for CUDA kernel launches.
           If negative or zero, the maximum hardware value will be used.
-        block_dim: The number of threads per block.
+        block_dim: The number of threads per block (always 1 for "cpu" devices).
     """
 
     init()
@@ -5476,6 +5671,9 @@ def launch(
     else:
         device = runtime.get_device(device)
 
+    if device == "cpu":
+        block_dim = 1
+
     # check function is a Kernel
     if not isinstance(kernel, Kernel):
         raise RuntimeError("Error launching kernel, can only launch functions decorated with @wp.kernel.")
@@ -5708,6 +5906,18 @@ def launch_tiled(*args, **kwargs):
             "Launch block dimension 'block_dim' argument should be passed via. keyword args for wp.launch_tiled()"
         )
 
+    if "device" in kwargs:
+        device = kwargs["device"]
+    else:
+        # todo: this doesn't consider the case where device
+        # is passed through positional args
+        device = None
+
+    # force the block_dim to 1 if running on "cpu"
+    device = runtime.get_device(device)
+    if device.is_cpu:
+        kwargs["block_dim"] = 1
+
     dim = kwargs["dim"]
     if not isinstance(dim, list):
         dim = list(dim) if isinstance(dim, tuple) else [dim]
@@ -5876,6 +6086,7 @@ def set_module_options(options: Dict[str, Any], module: Optional[Any] = None):
 
     * **mode**: The compilation mode to use, can be "debug", or "release", defaults to the value of ``warp.config.mode``.
     * **max_unroll**: The maximum fixed-size loop to unroll, defaults to the value of ``warp.config.max_unroll``.
+    * **block_dim**: The default number of threads to assign to each block
 
     Args:
 
@@ -5901,7 +6112,12 @@ def get_module_options(module: Optional[Any] = None) -> Dict[str, Any]:
     return get_module(m.__name__).options
 
 
-def capture_begin(device: Devicelike = None, stream=None, force_module_load=None, external=False):
+def capture_begin(
+    device: Devicelike = None,
+    stream: Optional[Stream] = None,
+    force_module_load: Optional[bool] = None,
+    external: bool = False,
+):
     """Begin capture of a CUDA graph
 
     Captures all subsequent kernel launches and memory operations on CUDA devices.
@@ -5968,16 +6184,15 @@ def capture_begin(device: Devicelike = None, stream=None, force_module_load=None
     runtime.captures[capture_id] = graph
 
 
-def capture_end(device: Devicelike = None, stream: Stream = None) -> Graph:
-    """Ends the capture of a CUDA graph
+def capture_end(device: Devicelike = None, stream: Optional[Stream] = None) -> Graph:
+    """End the capture of a CUDA graph.
 
     Args:
-
         device: The CUDA device where capture began
         stream: The CUDA stream where capture began
 
     Returns:
-        A Graph object that can be launched with :func:`~warp.capture_launch()`
+        A :class:`Graph` object that can be launched with :func:`~warp.capture_launch()`
     """
 
     if stream is not None:
@@ -6011,12 +6226,12 @@ def capture_end(device: Devicelike = None, stream: Stream = None) -> Graph:
     return graph
 
 
-def capture_launch(graph: Graph, stream: Stream = None):
+def capture_launch(graph: Graph, stream: Optional[Stream] = None):
     """Launch a previously captured CUDA graph
 
     Args:
-        graph: A Graph as returned by :func:`~warp.capture_end()`
-        stream: A Stream to launch the graph on (optional)
+        graph: A :class:`Graph` as returned by :func:`~warp.capture_end()`
+        stream: A :class:`Stream` to launch the graph on
     """
 
     if stream is not None:
@@ -6032,24 +6247,28 @@ def capture_launch(graph: Graph, stream: Stream = None):
 
 
 def copy(
-    dest: warp.array, src: warp.array, dest_offset: int = 0, src_offset: int = 0, count: int = 0, stream: Stream = None
+    dest: warp.array,
+    src: warp.array,
+    dest_offset: int = 0,
+    src_offset: int = 0,
+    count: int = 0,
+    stream: Optional[Stream] = None,
 ):
     """Copy array contents from `src` to `dest`.
 
     Args:
-        dest: Destination array, must be at least as big as source buffer
+        dest: Destination array, must be at least as large as source buffer
         src: Source array
         dest_offset: Element offset in the destination array
         src_offset: Element offset in the source array
         count: Number of array elements to copy (will copy all elements if set to 0)
-        stream: The stream on which to perform the copy (optional)
+        stream: The stream on which to perform the copy
 
     The stream, if specified, can be from any device.  If the stream is omitted, then Warp selects a stream based on the following rules:
     (1) If the destination array is on a CUDA device, use the current stream on the destination device.
     (2) Otherwise, if the source array is on a CUDA device, use the current stream on the source device.
 
     If neither source nor destination are on a CUDA device, no stream is used for the copy.
-
     """
 
     from warp.context import runtime
@@ -6274,8 +6493,8 @@ def type_str(t):
             return f"Transformation[{type_str(t._wp_scalar_type_)}]"
 
         raise TypeError("Invalid vector or matrix dimensions")
-    elif warp.codegen.get_type_origin(t) in (list, tuple):
-        args_repr = ", ".join(type_str(x) for x in warp.codegen.get_type_args(t))
+    elif get_origin(t) in (list, tuple):
+        args_repr = ", ".join(type_str(x) for x in get_args(t))
         return f"{t._name}[{args_repr}]"
     elif t is Ellipsis:
         return "..."