triton-windows 3.4.0.post20__cp312-cp312-win_amd64.whl → 3.5.0.post21__cp312-cp312-win_amd64.whl

triton/compiler/compiler.py

@@ -7,15 +7,15 @@ from ..backends.compiler import Language
 from ..backends.compiler import BaseBackend, GPUTarget
 from .. import __version__, knobs
 from ..runtime.autotuner import OutOfResources
-from ..runtime.cache import get_cache_manager, get_dump_manager, get_override_manager
+from ..runtime.cache import get_cache_manager, get_dump_manager, get_override_manager, get_cache_key
 from ..runtime.driver import driver
 from ..tools.disasm import get_sass
 from pathlib import Path
 import re
 import functools
 import os
-import sysconfig
 import time
+import copy
 
 # - ^\s*tt\.func\s+ : match the start of the string, any leading whitespace, the keyword func,
 #    and any following whitespace
@@ -64,12 +64,9 @@ class ASTSource:
                 assert isinstance(k, tuple)
                 self.constants[k] = v
         self.attrs = attrs or dict()
-        if isinstance(self.signature, str):
-            self.signature = {k: v.strip() for k, v in enumerate(self.signature.split(","))}
-        else:
-            for k in self.signature.keys():
-                if not isinstance(k, str):
-                    raise TypeError("Signature keys must be string")
+        for k in self.signature.keys():
+            if not isinstance(k, str):
+                raise TypeError("Signature keys must be string")
 
     def hash(self):
         sorted_sig = [v for k, v in sorted(self.signature.items())]
@@ -78,7 +75,7 @@ class ASTSource:
         key = f"{self.fn.cache_key}-{str(self.attrs)}-{sorted_sig}-{constants_key}"
         return hashlib.sha256(key.encode("utf-8")).hexdigest()
 
-    def make_ir(self, options, codegen_fns, module_map, context):
+    def make_ir(self, target: GPUTarget, options, codegen_fns, module_map, context):
         from .code_generator import ast_to_ttir
         return ast_to_ttir(self.fn, self, context=context, options=options, codegen_fns=codegen_fns,
                            module_map=module_map)
@@ -117,7 +114,7 @@ class IRSource:
     def hash(self):
         return hashlib.sha256(self.src.encode("utf-8")).hexdigest()
 
-    def make_ir(self, options, codegen_fns, module_map, context):
+    def make_ir(self, target: GPUTarget, options, codegen_fns, module_map, context):
         self.module.context = context
         return self.module
 
@@ -129,42 +126,6 @@ class IRSource:
         return dict()
 
 
-@functools.lru_cache()
-def triton_key():
-    import pkgutil
-    TRITON_PATH = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
-    contents = []
-    # frontend
-    with open(__file__, "rb") as f:
-        contents += [hashlib.sha256(f.read()).hexdigest()]
-    # compiler
-    path_prefixes = [
-        (os.path.join(TRITON_PATH, "compiler"), "triton.compiler."),
-        (os.path.join(TRITON_PATH, "backends"), "triton.backends."),
-    ]
-    for path, prefix in path_prefixes:
-        for lib in pkgutil.walk_packages([path], prefix=prefix):
-            with open(lib.module_finder.find_spec(lib.name).origin, "rb") as f:
-                contents += [hashlib.sha256(f.read()).hexdigest()]
-
-    # backend
-    libtriton_hash = hashlib.sha256()
-    ext = sysconfig.get_config_var("EXT_SUFFIX").split(".")[-1]
-    with open(os.path.join(TRITON_PATH, "_C", f"libtriton.{ext}"), "rb") as f:
-        while True:
-            chunk = f.read(1024**2)
-            if not chunk:
-                break
-            libtriton_hash.update(chunk)
-    contents.append(libtriton_hash.hexdigest())
-    # language
-    language_path = os.path.join(TRITON_PATH, 'language')
-    for lib in pkgutil.walk_packages([language_path], prefix="triton.language."):
-        with open(lib.module_finder.find_spec(lib.name).origin, "rb") as f:
-            contents += [hashlib.sha256(f.read()).hexdigest()]
-    return f'{__version__}' + '-'.join(contents)
-
-
 @functools.lru_cache()
 def max_shared_mem(device):
     return driver.active.utils.get_device_properties(device)["max_shared_mem"]
@@ -258,7 +219,7 @@ class CompileTimer:
         )
 
 
-def compile(src, target=None, options=None):
+def compile(src, target=None, options=None, _env_vars=None):
     compilation_listener = knobs.compilation.listener
     if compilation_listener:
         timer = CompileTimer()
@@ -277,8 +238,8 @@ def compile(src, target=None, options=None):
     extra_options = src.parse_options()
     options = backend.parse_options(dict(options or dict(), **extra_options))
     # create cache manager
-    env_vars = get_cache_invalidating_env_vars()
-    key = f"{triton_key()}-{src.hash()}-{backend.hash()}-{options.hash()}-{str(sorted(env_vars.items()))}"
+    env_vars = get_cache_invalidating_env_vars() if _env_vars is None else _env_vars
+    key = get_cache_key(src, backend, options, env_vars=env_vars)
     hash = hashlib.sha256(key.encode("utf-8")).hexdigest()
     fn_cache_manager = get_cache_manager(hash)
     # For dumping/overriding only hash the source as we want it to be independent of triton
@@ -336,7 +297,7 @@ def compile(src, target=None, options=None):
     codegen_fns = backend.get_codegen_implementation(options)
     module_map = backend.get_module_map()
     try:
-        module = src.make_ir(options, codegen_fns, module_map, context)
+        module = src.make_ir(target, options, codegen_fns, module_map, context)
     except Exception as e:
         filter_traceback(e)
         raise
@@ -371,6 +332,9 @@ def compile(src, target=None, options=None):
             metadata_group[ir_filename] = fn_cache_manager.put(next_module, ir_filename)
         if fn_dump_manager is not None:
             fn_dump_manager.put(next_module, ir_filename)
+            if ext == "cubin":
+                sass = get_sass(next_module)
+                fn_dump_manager.put(sass, file_name + ".sass")
         # use an env variable to parse ir from file
         if use_ir_loc == ext:
             ir_full_name = fn_cache_manager.get_file(ir_filename)
@@ -440,6 +404,10 @@ class AsmDict(dict):
         return value
 
 
+def _raise_error(err, *args, **kwargs):
+    raise copy.deepcopy(err)
+
+
 class CompiledKernel:
 
     def __init__(self, src, metadata_group, hash):
@@ -464,51 +432,66 @@ class CompiledKernel:
             file.suffix[1:]: file.read_bytes() if file.suffix[1:] == binary_ext else file.read_text()
             for file in asm_files
         })
+        self.metadata_group = metadata_group
         self.kernel = self.asm[binary_ext]
         # binaries are lazily initialized
         # because it involves doing runtime things
         # (e.g., checking amount of shared memory on current device)
         self.module = None
         self.function = None
+        self._run = None
 
     def _init_handles(self):
         if self.module is not None:
             return
+
+        def raise_(err):
+            # clone the exception object so that the one saved in the closure
+            # of the partial function below doesn't get assigned a stack trace
+            # after the subsequent raise. otherwise, the CompiledKernel instance
+            # saved in the (global) kernel cache will keep references to all the
+            # locals in the traceback via the exception instance in the closure.
+            cloned_err = copy.deepcopy(err)
+            self._run = functools.partial(_raise_error, cloned_err)
+            raise err
+
         device = driver.active.get_current_device()
         # create launcher
-        self.run = driver.active.launcher_cls(self.src, self.metadata)
+        self._run = driver.active.launcher_cls(self.src, self.metadata)
         # not enough shared memory to run the kernel
         max_shared = max_shared_mem(device)
         if self.metadata.shared > max_shared:
-            raise OutOfResources(self.metadata.shared, max_shared, "shared memory")
+            raise_(OutOfResources(self.metadata.shared, max_shared, "shared memory"))
         if hasattr(self.metadata, "tmem_size") and self.metadata.tmem_size is not None:
             # Use blackwell max tmem size for now, this should be moved in device properties
             max_tmem_size = 512  # tmem size in number of columns
             if self.metadata.tmem_size > max_tmem_size:
-                raise OutOfResources(self.metadata.tmem_size, max_tmem_size, "tensor memory")
+                raise_(OutOfResources(self.metadata.tmem_size, max_tmem_size, "tensor memory"))
+        if knobs.runtime.kernel_load_start_hook is not None:
+            knobs.runtime.kernel_load_start_hook(self.module, self.function, self.name, self.metadata_group, self.hash)
         # TODO: n_regs, n_spills should be metadata generated when calling `ptxas`
         self.module, self.function, self.n_regs, self.n_spills, self.n_max_threads = driver.active.utils.load_binary(
             self.name, self.kernel, self.metadata.shared, device)
         warp_size = driver.active.get_current_target().warp_size
         if self.metadata.num_warps * warp_size > self.n_max_threads:
-            raise OutOfResources(self.metadata.num_warps * warp_size, self.n_max_threads, "threads")
+            raise_(OutOfResources(self.metadata.num_warps * warp_size, self.n_max_threads, "threads"))
+        if knobs.runtime.kernel_load_end_hook is not None:
+            knobs.runtime.kernel_load_end_hook(self.module, self.function, self.name, self.metadata_group, self.hash)
 
-    def __getattribute__(self, name):
-        if name == 'run':
+    @property
+    def run(self):
+        if self._run is None:
             self._init_handles()
-        return super().__getattribute__(name)
+        return self._run
 
     def launch_metadata(self, grid, stream, *args):
         if knobs.runtime.launch_enter_hook is None:
             return None
+        self._init_handles()
         ret = LazyDict({"name": self.name, "function": self.function, "stream": stream})
         if not isinstance(self.src, ASTSource) or self.src.fn.launch_metadata is None:
             return ret
-        arg_dict = {}
-        arg_idx = 0
-        for i, arg_name in enumerate(self.src.fn.arg_names):
-            arg_dict[arg_name] = args[arg_idx]
-            arg_idx += 1
+        arg_dict = {name: arg for name, arg in zip(self.src.fn.arg_names, args)}
         ret.add(self.src.fn.launch_metadata, (grid, self.metadata, arg_dict))
         return ret
 

triton/experimental/gluon/__init__.py

@@ -1,4 +1,5 @@
 from . import nvidia
-from ._runtime import jit
+from ._runtime import constexpr_function, jit
+from triton.language.core import must_use_result
 
-__all__ = ["jit", "nvidia"]
+__all__ = ["constexpr_function", "jit", "must_use_result", "nvidia"]

triton/experimental/gluon/_runtime.py

@@ -1,13 +1,14 @@
 from __future__ import annotations
-import triton
 from triton.compiler.compiler import ASTSource
 from triton.backends.compiler import Language
-from triton.runtime.jit import JITFunction
+from triton.runtime.jit import JITFunction, constexpr_function
 from typing import TypeVar, Optional, Callable, Iterable, Union
 from triton._C.libtriton import ir
 
 T = TypeVar("T")
 
+__all__ = ["constexpr_function", "jit"]
+
 
 class GluonASTSource(ASTSource):
 
@@ -16,7 +17,7 @@ class GluonASTSource(ASTSource):
         self.language = Language.GLUON
         self.ext = "ttgir"
 
-    def make_ir(self, options, codegen_fns, module_map, context):
+    def make_ir(self, target, options, codegen_fns, module_map, context):
         from triton.compiler.compiler import make_backend
         from triton.compiler.code_generator import ast_to_ttir
 
@@ -24,14 +25,16 @@ class GluonASTSource(ASTSource):
         module = builder.create_module()
 
         # Assign module attributes eagerly, as they are needed to verify layouts
-        target = triton.runtime.driver.active.get_current_target()
         backend = make_backend(target)
         target = backend.get_target_name(options)
+
         module.set_attr("ttg.target", builder.get_string_attr(target))
         module.set_attr("ttg.num-warps", builder.get_int32_attr(options.num_warps))
         module.set_attr("ttg.num-ctas", builder.get_int32_attr(options.num_ctas))
-        module.set_attr("ttg.threads-per-warp", builder.get_int32_attr(32))
-        if options.maxnreg is not None:
+        module.set_attr("ttg.threads-per-warp", builder.get_int32_attr(options.warp_size))
+
+        is_cuda = options.backend_name == "cuda"
+        if is_cuda and options.maxnreg is not None:
             module.set_attr("ttg.maxnreg", builder.get_int32_attr(options.maxnreg))
 
         module = ast_to_ttir(self.fn, self, context=context, options=options, codegen_fns=codegen_fns,

triton/experimental/gluon/language/__init__.py

@@ -1,18 +1,119 @@
-from ._core import *  # NOQA: F403
-from ._core import __all__ as __core_all
-from ._layouts import *  # NOQA: F403
-from ._layouts import __all__ as __layouts_all
-from ._math import *  # NOQA: F403
-from ._math import __all__ as __math_all
-from ._standard import *  # NOQA: F403
-from ._standard import __all__ as __standard_all
+from ._core import (
+    base_value,
+    base_type,
+    block_type,
+    broadcast,
+    constexpr,
+    dtype,
+    void,
+    int1,
+    int8,
+    int16,
+    int32,
+    int64,
+    uint8,
+    uint16,
+    uint32,
+    uint64,
+    float8e5,
+    float8e5b16,
+    float8e4nv,
+    float8e4b8,
+    float8e4b15,
+    float16,
+    bfloat16,
+    float32,
+    float64,
+    pointer_type,
+    shared_memory_descriptor,
+    tensor,
+    tuple,
+    tuple_type,
+    _unwrap_if_constexpr,
+    # API Functions
+    allocate_shared_memory,
+    arange,
+    associative_scan,
+    atomic_add,
+    atomic_and,
+    atomic_cas,
+    atomic_max,
+    atomic_min,
+    atomic_or,
+    atomic_xchg,
+    atomic_xor,
+    convert_layout,
+    device_assert,
+    expand_dims,
+    full,
+    histogram,
+    inline_asm_elementwise,
+    join,
+    load,
+    map_elementwise,
+    max_constancy,
+    max_contiguous,
+    maximum,
+    minimum,
+    multiple_of,
+    num_programs,
+    permute,
+    program_id,
+    reduce,
+    reshape,
+    set_auto_layout,
+    split,
+    static_assert,
+    static_print,
+    static_range,
+    store,
+    thread_barrier,
+    to_tensor,
+    warp_specialize,
+    where,
+)
+from ._layouts import (
+    AutoLayout,
+    BlockedLayout,
+    SliceLayout,
+    DistributedLinearLayout,
+    DotOperandLayout,
+    NVMMADistributedLayout,
+    NVMMASharedLayout,
+    SwizzledSharedLayout,
+    PaddedSharedLayout,
+)
+from ._math import (
+    umulhi,
+    exp,
+    exp2,
+    fma,
+    log,
+    log2,
+    cos,
+    rsqrt,
+    sin,
+    sqrt,
+    sqrt_rn,
+    abs,
+    fdiv,
+    div_rn,
+    erf,
+    floor,
+    ceil,
+)
+from ._standard import (
+    cdiv,
+    full_like,
+    max,
+    min,
+    reduce_or,
+    sum,
+    xor_sum,
+    zeros,
+    zeros_like,
+)
 
 from . import nvidia
-
-__all__ = [
-    *__core_all,
-    *__layouts_all,
-    *__math_all,
-    *__standard_all,
-    "nvidia",
-]
+from . import amd
+from . import extra