triton-windows 3.4.0.post20__cp313-cp313-win_amd64.whl → 3.5.0.post21__cp313-cp313-win_amd64.whl

triton/experimental/gluon/language/nvidia/hopper/__init__.py

@@ -1,11 +1,132 @@
-from . import mbarrier
-from . import tma
+from __future__ import annotations
+from triton.compiler.code_generator import unflatten_ir_values
+from ..ampere import async_copy
+from . import mbarrier, tma
 from ... import _core
 
-__all__ = ["fence_async_shared", "mbarrier", "tma"]
+from typing import List, Tuple, TYPE_CHECKING
+if TYPE_CHECKING:
+    from triton._C.libtriton import ir
+
+__all__ = ["async_copy", "fence_async_shared", "mbarrier", "tma", "warpgroup_mma", "warpgroup_mma_wait"]
 
 
 @_core.builtin
 def fence_async_shared(cluster=False, _semantic=None):
+    """
+    Issue a fence to complete asynchronous shared memory operations.
+
+    Args:
+        cluster (bool): Whether to fence across cluster. Defaults to False.
+    """
     cluster = _core._unwrap_if_constexpr(cluster)
     _semantic.builder.create_fence_async_shared(cluster)
+
+
+class warpgroup_mma_accumulator_type(_core.base_type):
+    tensor_type: _core.dtype
+
+    def __init__(self, tensor_type: _core.dtype):
+        self.tensor_type = tensor_type
+
+    def __str__(self) -> str:
+        return f"warpgroup_mma_accumulator<{self.tensor_type}>"
+
+    def _unflatten_ir(self, handles: List[ir.value], cursor: int) -> Tuple[warpgroup_mma_accumulator, int]:
+        return warpgroup_mma_accumulator(handles[cursor], self.tensor_type), cursor + 1
+
+    def _flatten_ir_types(self, builder: ir.builder, out: List[ir.type]) -> None:
+        self.tensor_type._flatten_ir_types(builder, out)
+
+    def __eq__(self, other) -> bool:
+        return type(self) is type(other) and self.tensor_type == other.tensor_type
+
+    def mangle(self) -> str:
+        return f"FT{self.tensor_type.mangle()}FT"
+
+
+class warpgroup_mma_accumulator(_core.base_value):
+    handle: ir.value
+    type: warpgroup_mma_accumulator_type
+
+    def __init__(self, handle, tensor_type: _core.dtype):
+        self.handle = handle
+        self.type = warpgroup_mma_accumulator_type(tensor_type)
+
+    def _flatten_ir(self, handles: List[ir.value]) -> None:
+        handles.append(self.handle)
+
+
+@_core.builtin
+def warpgroup_mma_init(value, _semantic):
+    assert isinstance(value, _core.tensor)
+    return warpgroup_mma_accumulator(value.handle, value.type)
+
+
+@_core.builtin
+def warpgroup_mma(a, b, acc, *, use_acc=True, precision=None, max_num_imprecise_acc=None, is_async=False,
+                  _semantic=None):
+    """
+    Perform warpgroup MMA (Tensor Core) operations.
+    acc = a * b + (acc if use_acc else 0)
+
+    Args:
+        a (tensor or shared_memory_descriptor): Left hand side operand.
+        b (shared_memory_descriptor): Right hand side operand.
+        acc (tensor): Accumulator tensor.
+        use_acc (bool): Whether to use the initial value of the accumulator. Defaults to True.
+        precision (str, optional): Dot input precision. Defaults to builder default.
+        max_num_imprecise_acc (int): Max imprecise accumulations. Used for fp8 -> fp32 dot. Determines how many accumulation are done in limited precision. Defaults to None, which means no upcasting is done.
+        is_async (bool): Whether operation is asynchronous. Defaults to False.
+
+    Returns:
+        tensor or warpgroup_mma_accumulator: Returns the result if synchronous, or a token to load the value once computed if asynchronous.
+    """
+    use_acc = _semantic.to_tensor(use_acc)
+
+    if precision is None:
+        precision = _semantic.builder.options.default_dot_input_precision
+
+    precision = _semantic._str_to_dot_input_precision(precision)
+
+    K = a.type.shape[-1]
+    if max_num_imprecise_acc is None:
+        if a.dtype.is_fp8() and b.dtype.is_fp8():
+            max_num_imprecise_acc = _semantic.builder.options.max_num_imprecise_acc_default
+        else:
+            max_num_imprecise_acc = 0
+    else:
+        if a.dtype.is_fp8() and b.dtype.is_fp8() and max_num_imprecise_acc > K:
+            raise ValueError(f"max_num_imprecise_acc ({max_num_imprecise_acc}) must be <= K ({K})")
+
+    max_num_imprecise_acc = _core._unwrap_if_constexpr(max_num_imprecise_acc)
+    is_async = _core._unwrap_if_constexpr(is_async)
+
+    handle = _semantic.builder.create_warpgroup_mma(a.handle, b.handle, acc.handle, use_acc.handle, precision,
+                                                    max_num_imprecise_acc, is_async)
+    tensor_ty = acc.type.tensor_type if isinstance(acc, warpgroup_mma_accumulator) else acc.type
+    if is_async:
+        return warpgroup_mma_accumulator(handle, tensor_ty)
+    else:
+        return _core.tensor(handle, tensor_ty)
+
+
+@_core.builtin
+def warpgroup_mma_wait(num_outstanding=0, deps=None, _semantic=None):
+    """
+    Wait until `num_outstanding` or less warpgroup MMA operations are in-flight.
+
+    Args:
+        num_outstanding (int): Number of outstanding warpgroup MMA operations to wait for. Defaults to 0.
+        deps (Sequence[tensor]): List of dependencies that need to be kept alive while the mma is unfinished.
+    """
+    if deps is None:
+        raise ValueError("warpgroup_mma_wait deps must be given")
+    deps_handles = [x.handle for x in deps] if deps is not None else []
+    num_outstanding = _core._unwrap_if_constexpr(num_outstanding)
+    results = _semantic.builder.create_warpgroup_mma_wait(deps_handles, num_outstanding)
+    result_types = [dep.type.tensor_type if isinstance(dep, warpgroup_mma_accumulator) else dep.type for dep in deps]
+    results = unflatten_ir_values(results, result_types)
+    if len(deps) == 1:
+        return next(results)
+    return tuple(results)

triton/experimental/gluon/language/nvidia/hopper/mbarrier.py

@@ -1,51 +1,34 @@
-from triton.experimental.gluon.language._layouts import SwizzledSharedLayout
-from triton.experimental.gluon.language._core import builtin, _unwrap_if_constexpr
+from ..ampere.mbarrier import MBarrierLayout, init, invalidate, wait
+from ..._core import _unwrap_if_constexpr, builtin
 
-__all__ = ["MBarrierLayout", "init", "invalidate", "expect", "wait", "arrive"]
-
-
-class MBarrierLayout(SwizzledSharedLayout):
-
-    def __init__(self, ctas_per_cga: int = 1, cta_split_num: int = 1):
-        super().__init__(
-            vec=1,
-            per_phase=1,
-            max_phase=1,
-            order=[0],
-            ctas_per_cga=[ctas_per_cga],
-            cta_split_num=[cta_split_num],
-            cta_order=[0],
-        )
-
-
-@builtin
-def init(mbarrier, count, _semantic=None):
-    count = _unwrap_if_constexpr(count)
-    _semantic.builder.create_mbarrier_init(mbarrier.handle, count)
-
-
-@builtin
-def invalidate(mbarrier, _semantic=None):
-    _semantic.builder.create_mbarrier_inval(mbarrier.handle)
+__all__ = ["arrive", "expect", "init", "invalidate", "MBarrierLayout", "wait"]
 
 
 @builtin
 def expect(mbarrier, bytes, pred=True, _semantic=None):
+    """
+    Expect a specific number of bytes being copied. When they are copied, the barrier is signaled.
+
+    Args:
+        mbarrier (shared_memory_descriptor): Barrier that will be signaled when the operation is complete.
+        bytes (int): Expected byte count.
+        pred (bool): Scalar predicate. Operation is skipped if predicate is False. Defaults to True.
+    """
     bytes = _unwrap_if_constexpr(bytes)
     pred = _semantic.to_tensor(pred)
     _semantic.builder.create_mbarrier_expect(mbarrier.handle, bytes, pred.handle)
 
 
 @builtin
-def wait(mbarrier, phase, pred=True, deps=(), _semantic=None):
-    phase = _semantic.to_tensor(phase)
-    pred = _semantic.to_tensor(pred)
-    deps = [x.handle for x in deps]
-    _semantic.builder.create_mbarrier_wait(mbarrier.handle, phase.handle, pred.handle, deps)
-
-
-@builtin
-def arrive(mbarrier, count, pred=True, _semantic=None):
+def arrive(mbarrier, *, count=1, pred=True, _semantic=None):
+    """
+    Arrive at an mbarrier with a specified count.
+
+    Args:
+        mbarrier (shared_memory_descriptor): Barrier to be signalled.
+        count (int): Count to arrive with. Defaults to 1.
+        pred (bool): Scalar predicate. Operation is skipped if predicate is False. Defaults to True.
+    """
     count = _unwrap_if_constexpr(count)
     pred = _semantic.to_tensor(pred)
     _semantic.builder.create_mbarrier_arrive(mbarrier.handle, count, pred.handle)

triton/experimental/gluon/language/nvidia/hopper/tma.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 from typing import List, Tuple, TYPE_CHECKING
 from dataclasses import dataclass
+from triton.language.core import base_type, base_value
 import triton.experimental.gluon.language._core as ttgl
 from triton.experimental.gluon.language._layouts import NVMMASharedLayout
 from triton.experimental.gluon.language._core import builtin, _unwrap_if_constexpr
@@ -12,7 +13,7 @@ __all__ = ["async_copy_global_to_shared", "async_copy_shared_to_global", "store_
 
 
 @dataclass(eq=True)
-class tensor_descriptor_type:
+class tensor_descriptor_type(base_type):
     block_type: ttgl.block_type
     shape_type: ttgl.tuple_type
     strides_type: ttgl.tuple_type
@@ -41,10 +42,10 @@ class tensor_descriptor_type:
         self.strides_type._flatten_ir_types(builder, out)
 
     def mangle(self) -> str:
-        return f"TD{self.block_type.mangle}_{self.layout.mangle()}TD"
+        return f"TD{self.block_type.mangle()}_{self.layout.mangle()}TD"
 
 
-class tensor_descriptor:
+class tensor_descriptor(base_value):
 
     def __init__(self, handle, shape: List[ttgl.tensor], strides: List[ttgl.tensor], block_type: ttgl.block_type,
                  layout: NVMMASharedLayout):

triton/experimental/gluon/nvidia/hopper.py

@@ -13,6 +13,7 @@ class TensorDescriptor:
     strides: List[int]
     block_shape: List[int]
     layout: NVMMASharedLayout
+    padding: str = "zero"
 
     def __post_init__(self):
         rank = len(self.shape)
@@ -28,13 +29,17 @@ class TensorDescriptor:
             assert (stride * elem_bytes) % 16 == 0, "strides must be 16-byte aligned"
         assert self.strides[-1] == 1, "Last dimension must be contiguous"
         assert isinstance(self.layout, NVMMASharedLayout), "Layout must be NVMMASharedLayout"
+        assert self.padding == "zero" or self.padding == "nan", "Illegal value for padding"
+        if self.padding == "nan":
+            assert self.base.dtype.is_floating_point, "Padding option `nan` is only supported for floating point tensors"
 
     @staticmethod
-    def from_tensor(tensor: Any, block_shape: List[int], layout: NVMMASharedLayout):
+    def from_tensor(tensor: Any, block_shape: List[int], layout: NVMMASharedLayout, padding="zero"):
         return TensorDescriptor(
             tensor,
             tensor.shape,
             tensor.stride(),
             block_shape,
             layout,
+            padding,
         )

triton/knobs.py

@@ -1,15 +1,19 @@
 from __future__ import annotations
 
+import functools
 import importlib
 import os
 import re
 import subprocess
 import sysconfig
+import warnings
 
 from dataclasses import dataclass
 from contextlib import contextmanager
 from typing import cast, Any, Callable, Generator, Generic, Optional, Protocol, Type, TypeVar, TypedDict, TYPE_CHECKING, Union
 
+from triton._C.libtriton import getenv, getenv_bool  # type: ignore
+
 if TYPE_CHECKING:
     from .runtime.cache import CacheManager, RemoteCacheBackend
     from .runtime.jit import JitFunctionInfo, KernelParam
@@ -25,11 +29,6 @@ env = Env()
 propagate_env: bool = True
 
 
-def getenv(key: str) -> Optional[str]:
-    res = os.getenv(key)
-    return res.strip() if res is not None else res
-
-
 def setenv(key: str, value: Optional[str]) -> None:
     if not propagate_env:
         return
@@ -62,32 +61,25 @@ def toenv(val: Any) -> Union[None, tuple[Optional[str]]]:
 SetType = TypeVar("SetType")
 GetType = TypeVar("GetType")
 
+_NOTHING = object()
+
 
 class env_base(Generic[SetType, GetType]):
 
-    def __init__(self, key: str, default: Union[SetType, Callable[[], SetType]]) -> None:
+    def __init__(self, key: str) -> None:
         self.key = key
-        self.default: Callable[[], SetType] = default if callable(default) else lambda: default
 
     def __set_name__(self, objclass: Type[object], name: str) -> None:
         self.name = name
 
     def __get__(self, obj: Optional[object], objclass: Optional[Type[object]]) -> GetType:
-        if obj is None:
-            raise AttributeError(f"Cannot access {type(self)} on non-instance")
-
-        if self.name in obj.__dict__:
-            return self.transform(obj.__dict__[self.name])
-        else:
+        py_val = obj.__dict__.get(self.name, _NOTHING)
+        if py_val is _NOTHING:
             return self.get()
-
-    @property
-    def env_val(self) -> str | None:
-        return getenv(self.key)
+        return self.transform(py_val)
 
     def get(self) -> GetType:
-        env = self.env_val
-        return self.transform(self.default() if env is None else self.from_env(env))
+        raise NotImplementedError()
 
     def __set__(self, obj: object, value: Union[SetType, Env]) -> None:
         if isinstance(value, Env):
@@ -105,54 +97,70 @@ class env_base(Generic[SetType, GetType]):
         # if GetType != SetType.
         return cast(GetType, val)
 
-    def from_env(self, val: str) -> SetType:
-        raise NotImplementedError()
-
 
 class env_str(env_base[str, str]):
 
-    def from_env(self, val: str) -> str:
-        return val
+    def __init__(self, key: str, default: str):
+        super().__init__(key)
+        self.default = default
+
+    def get(self) -> str:
+        return getenv(self.key, self.default)
+
+
+class env_str_callable_default(env_base[str, str]):
+
+    def __init__(self, key: str, default_factory: Callable[[], str]):
+        super().__init__(key)
+        self.default_factory = default_factory
+
+    def get(self) -> str:
+        env_val = getenv(self.key)
+        if env_val is None:
+            return self.default_factory()
+        return env_val
 
 
 class env_bool(env_base[bool, bool]):
 
-    def __init__(self, key: str, default: Union[bool, Callable[[], bool]] = False) -> None:
-        super().__init__(key, default)
+    def __init__(self, key: str, default: bool = False) -> None:
+        super().__init__(key)
+        self.default = default
 
-    def from_env(self, val: str) -> bool:
-        return val.lower() in ("1", "true", "yes", "on", "y")
+    def get(self) -> bool:
+        return getenv_bool(self.key, self.default)
 
 
 class env_int(env_base[int, int]):
 
-    def __init__(self, key: str, default: Union[int, Callable[[], int]] = 0) -> None:
-        super().__init__(key, default)
+    def __init__(self, key: str, default: int = 0) -> None:
+        super().__init__(key)
+        self.default = default
 
-    def from_env(self, val: str) -> int:
+    def get(self) -> int:
+        val = getenv(self.key)
+        if val is None:
+            return self.default
         try:
             return int(val)
         except ValueError as exc:
             raise RuntimeError(f"Unable to use {self.key}={val}: expected int") from exc
 
 
-class env_opt_base(Generic[GetType, SetType], env_base[Optional[GetType], Optional[SetType]]):
-
-    def __init__(self, key: str) -> None:
-        super().__init__(key, None)
-
-
 ClassType = TypeVar("ClassType")
 
 
-class env_class(Generic[ClassType], env_opt_base[Type[ClassType], Type[ClassType]]):
+class env_class(Generic[ClassType], env_base[Optional[Type[ClassType]], Optional[Type[ClassType]]]):
 
     def __init__(self, key: str, type: str) -> None:
         super().__init__(key)
         # We can't pass the type directly to avoid import cycles
         self.type = type
 
-    def from_env(self, val: str) -> Type[ClassType]:
+    def get(self) -> Optional[Type[ClassType]]:
+        val = getenv(self.key)
+        if val is None:
+            return None
         comps = val.split(":", 1)
         if len(comps) != 2:
             raise RuntimeError(f"Unable to read {self.key}: '{val}' isn't of the form MODULE:CLASS")
@@ -170,16 +178,15 @@ class NvidiaTool:
     version: str
 
     @staticmethod
+    @functools.lru_cache
     def from_path(path: str) -> Optional[NvidiaTool]:
         try:
             result = subprocess.check_output([path, "--version"], stderr=subprocess.STDOUT)
-            if result is None:
-                return None
             version = re.search(r".*release (\d+\.\d+).*", result.decode("utf-8"), flags=re.MULTILINE)
             if version is None:
                 return None
             return NvidiaTool(path, version.group(1))
-        except subprocess.CalledProcessError:
+        except (subprocess.CalledProcessError, FileNotFoundError):
             return None
 
 
@@ -202,6 +209,7 @@ def find_nvidia_tool(binary: str) -> str:
             if os.access(path, os.X_OK):
                 return path
 
+    warnings.warn(f"Failed to find {binary}")
     return ""
 
 
@@ -210,34 +218,38 @@ class env_nvidia_tool(env_base[str, NvidiaTool]):
     def __init__(self, binary: str) -> None:
         binary += sysconfig.get_config_var("EXE")
         self.binary = binary
-        super().__init__(f"TRITON_{binary.upper()}_PATH", lambda: find_nvidia_tool(self.binary))
+        self.default_path = find_nvidia_tool(binary)
+        super().__init__(f"TRITON_{binary.upper()}_PATH")
+
+    def get(self) -> NvidiaTool:
+        return self.transform(getenv(self.key))
 
     def transform(self, path: str) -> NvidiaTool:
-        paths = [
-            path,
-            # We still add default as fallback in case the pointed binary isn't
-            # accessible.
-            self.default(),
-        ]
+        # We still add default as fallback in case the pointed binary isn't
+        # accessible.
+        if path is not None:
+            paths = [path, self.default_path]
+        else:
+            paths = [self.default_path]
+
         for path in paths:
-            if not path or not os.access(path, os.X_OK):
-                continue
             if tool := NvidiaTool.from_path(path):
                 return tool
 
         raise RuntimeError(f"Cannot find {self.binary}")
 
-    def from_env(self, val: str) -> str:
-        return val
-
 
 # Separate classes so that types are correct
-class env_opt_str(env_opt_base[str, str], env_str):
-    pass
+class env_opt_str(env_base[Optional[str], Optional[str]]):
+
+    def get(self) -> Optional[str]:
+        return getenv(self.key)
 
 
-class env_opt_bool(env_opt_base[bool, bool], env_bool):
-    pass
+class env_opt_bool(env_base):
+
+    def get(self) -> Optional[str]:
+        return getenv_bool(self.key, None)
 
 
 @dataclass(frozen=True)
@@ -305,7 +317,7 @@ class base_knobs:
     @contextmanager
     def scope(self) -> Generator[None, None, None]:
         try:
-            initial_env = {knob.key: knob.env_val for knob in self.knob_descriptors.values()}
+            initial_env = {knob.key: getenv(knob.key) for knob in self.knob_descriptors.values()}
             orig = dict(self.__dict__)
             yield
         finally:
@@ -350,11 +362,11 @@ cache: cache_knobs
 
 
 class cache_knobs(base_knobs):
-    home_dir: env_str = env_str("TRITON_HOME", lambda: os.path.expanduser("~/"))
+    home_dir: env_str = env_str("TRITON_HOME", os.path.expanduser("~/"))
 
-    dump_dir: env_str = env_str("TRITON_DUMP_DIR", lambda: cache.get_triton_dir("dump"))
-    override_dir: env_str = env_str("TRITON_OVERRIDE_DIR", lambda: cache.get_triton_dir("override"))
-    dir: env_str = env_str("TRITON_CACHE_DIR", lambda: cache.get_triton_dir("cache"))
+    dump_dir = env_str_callable_default("TRITON_DUMP_DIR", lambda: cache.get_triton_dir("dump"))
+    override_dir = env_str_callable_default("TRITON_OVERRIDE_DIR", lambda: cache.get_triton_dir("override"))
+    dir = env_str_callable_default("TRITON_CACHE_DIR", lambda: cache.get_triton_dir("cache"))
 
     manager_class: env_class[CacheManager] = env_class("TRITON_CACHE_MANAGER", "CacheManager")
     remote_manager_class: env_class[RemoteCacheBackend] = env_class("TRITON_REMOTE_CACHE_BACKEND", "RemoteCacheBackend")
@@ -374,6 +386,7 @@ class compilation_knobs(base_knobs):
     disable_line_info: env_bool = env_bool("TRITON_DISABLE_LINE_INFO")
     front_end_debugging: env_bool = env_bool("TRITON_FRONT_END_DEBUGGING")
     allow_non_constexpr_globals: env_bool = env_bool("TRITON_ALLOW_NON_CONSTEXPR_GLOBALS")
+    enable_experimental_consan: env_bool = env_bool("TRITON_ENABLE_EXPERIMENTAL_CONSAN")
     listener: Union[CompilationListener, None] = None
 
 
@@ -383,11 +396,53 @@ class autotuning_knobs(base_knobs):
 
 
 class LaunchHook(Protocol):
+    """Hook invoked before and after kernel launching
+    """
 
     def __call__(self, metadata: LazyDict) -> None:
         ...
 
 
+class InitHandleHook(Protocol):
+    """Hook invoked around kernel binary/module loading.
+    module/function can be None for the *start* hook (before loading).
+    """
+
+    def __call__(
+        self,
+        module: Optional[object],
+        function: Optional[Callable],
+        name: str,
+        metadata_group: dict[str, str],
+        hash: str,
+    ) -> None:
+        ...
+
+
+F = TypeVar("F", bound=Callable)
+
+
+class HookChain(Generic[F]):
+    """A chain of hooks of the same type F to be called in order.
+    """
+
+    def __init__(self, reversed: bool = False):
+        self.calls: list[F] = []
+        self.reversed = reversed
+
+    def add(self, func: F) -> None:
+        if func not in self.calls:
+            self.calls.append(func)
+
+    def remove(self, func: F) -> None:
+        if func in self.calls:
+            self.calls.remove(func)
+
+    def __call__(self, *args, **kwargs):
+        for call in self.calls if not self.reversed else reversed(self.calls):
+            call(*args, **kwargs)
+
+
 # This is of the form [attr_name, attr_val]
 # TODO: Use tuple instead of list for better typing.
 KernelAttr = list[Union[str, int]]
@@ -418,11 +473,15 @@ class JITHook(Protocol):
 
 class runtime_knobs(base_knobs):
     interpret: env_bool = env_bool("TRITON_INTERPRET")
-    debug: env_bool = env_bool("TRITON_DEBUG")
+    # debug is on critical path for kernel launches
+    # avoid repeated reads from env-var by calling get directly
+    debug: bool = env_bool("TRITON_DEBUG").get()
     override_arch: env_opt_str = env_opt_str("TRITON_OVERRIDE_ARCH")
 
-    launch_enter_hook: Optional[LaunchHook] = None
-    launch_exit_hook: Optional[LaunchHook] = None
+    launch_enter_hook: HookChain[LaunchHook] = HookChain()
+    launch_exit_hook: HookChain[LaunchHook] = HookChain(reversed=True)
+    kernel_load_start_hook: HookChain[InitHandleHook] = HookChain()
+    kernel_load_end_hook: HookChain[InitHandleHook] = HookChain(reversed=True)
 
     # Hook for inspecting compiled functions and modules
     jit_cache_hook: Optional[JITHook] = None
@@ -444,6 +503,7 @@ class nvidia_knobs(base_knobs):
     dump_nvptx: env_bool = env_bool("NVPTX_ENABLE_DUMP")
     disable_ptxas_opt: env_bool = env_bool("DISABLE_PTXAS_OPT")
     mock_ptx_version: env_opt_str = env_opt_str("TRITON_MOCK_PTX_VERSION")
+    dump_ptxas_log: env_bool = env_bool("TRITON_DUMP_PTXAS_LOG")
 
     libdevice_path: env_opt_str = env_opt_str("TRITON_LIBDEVICE_PATH")
     libcuda_path: env_opt_str = env_opt_str("TRITON_LIBCUDA_PATH")
@@ -451,9 +511,10 @@ class nvidia_knobs(base_knobs):
 
 class amd_knobs(base_knobs):
     use_buffer_ops: env_bool = env_bool("AMDGCN_USE_BUFFER_OPS")
+    # Note: This requires use_buffer_ops be true to have any effect
+    use_buffer_atomics: env_bool = env_bool("AMDGCN_USE_BUFFER_ATOMICS", True)
     dump_amdgcn: env_bool = env_bool("AMDGCN_ENABLE_DUMP")
     libhip_path: env_opt_str = env_opt_str("TRITON_LIBHIP_PATH")
-    lld_path: env_opt_str = env_opt_str("TRITON_HIP_LLD_PATH")
 
     # We use strs so that we can have a default value based on other runtime info
     use_block_pingpong: env_opt_bool = env_opt_bool("TRITON_HIP_USE_BLOCK_PINGPONG")
@@ -479,3 +540,7 @@ language = language_knobs()
 nvidia = nvidia_knobs()
 amd = amd_knobs()
 proton = proton_knobs()
+
+
+def refresh_knobs():
+    runtime.debug = env_bool("TRITON_DEBUG").get()