triton-windows 3.3.0a0.post16__cp311-cp311-win_amd64.whl → 3.3.0a0.post18__cp311-cp311-win_amd64.whl

triton/backends/amd/compiler.py

@@ -28,6 +28,10 @@ class HIPOptions:
     waves_per_eu: int = 1
     num_stages: int = 2
     num_ctas: int = 1
+    num_buffers_warp_spec: int = 0
+    num_consumer_groups: int = 0
+    reg_dec_producer: int = 0
+    reg_inc_consumer: int = 0
     extern_libs: dict = None
     cluster_dims: tuple = (1, 1, 1)
     debug: bool = False

triton/backends/nvidia/compiler.py

@@ -141,6 +141,10 @@ class CUDAOptions:
     num_warps: int = 4
     num_ctas: int = 1
     num_stages: int = 3
+    num_buffers_warp_spec: int = 0
+    num_consumer_groups: int = 0
+    reg_dec_producer: int = 0
+    reg_inc_consumer: int = 0
     # maxnreg corresponds to the ptx parameter .maxnreg, which controls the
     # maximum number of 32-bit registers used by one thread.
     maxnreg: Optional[int] = None
@@ -285,16 +289,29 @@ class CUDABackend(BaseBackend):
             passes.ttgpuir.add_optimize_accumulator_init(pm)
             passes.common.add_canonicalizer(pm)
             passes.ttgpuir.add_combine_tensor_select_and_if(pm)
+            passes.ttgpuir.add_ws_task_partition(pm, opt.num_consumer_groups)
+            passes.ttgpuir.add_taskid_propagate(pm, opt.num_consumer_groups)
+            passes.ttgpuir.add_ws_data_partition(pm, opt.num_consumer_groups)
+            passes.ttgpuir.add_ws_code_partition(pm, opt.num_buffers_warp_spec, opt.num_consumer_groups,
+                                                 opt.reg_dec_producer, opt.reg_inc_consumer)
             passes.ttgpuir.add_pipeline(pm, opt.num_stages, dump_enabled)
+            passes.ttgpuir.add_ping_pong_sync(pm, opt.num_consumer_groups)
+            passes.ttgpuir.add_ws_lowering(pm, opt.num_consumer_groups)
         elif capability // 10 >= 10:
             passes.ttgpuir.add_fuse_nested_loops(pm)
             passes.common.add_canonicalizer(pm)
             passes.common.add_licm(pm)
             passes.ttgpuir.add_optimize_accumulator_init(pm)
+            passes.ttgpuir.add_ws_task_partition(pm, opt.num_consumer_groups)
+            passes.ttgpuir.add_taskid_propagate(pm, opt.num_consumer_groups)
+            passes.ttgpuir.add_ws_data_partition(pm, opt.num_consumer_groups)
+            passes.ttgpuir.add_ws_code_partition(pm, opt.num_buffers_warp_spec, opt.num_consumer_groups,
+                                                 opt.reg_dec_producer, opt.reg_inc_consumer)
             passes.ttgpuir.add_pipeline(pm, opt.num_stages, dump_enabled)
             passes.ttgpuir.add_combine_tensor_select_and_if(pm)
             nvidia.passes.ttnvgpuir.add_promote_lhs_to_tmem(pm)
             nvidia.passes.ttnvgpuir.add_keep_acc_in_tmem(pm)
+            passes.ttgpuir.add_ws_lowering(pm, opt.num_consumer_groups)
             passes.common.add_canonicalizer(pm)
         else:
             passes.common.add_licm(pm)
@@ -310,6 +327,8 @@ class CUDABackend(BaseBackend):
             nvidia.passes.ttnvgpuir.add_fence_insertion(pm)
             nvidia.passes.ttnvgpuir.add_tma_lowering(pm)
         passes.common.add_canonicalizer(pm)
+        if capability // 10 >= 9:
+            passes.ttgpuir.add_ws_canonicalization(pm, opt.num_consumer_groups)
         pm.run(mod)
         metadata["cluster_dims"] = (cluster_info.clusterDimX, cluster_info.clusterDimY, cluster_info.clusterDimZ)
         return mod
@@ -416,7 +435,9 @@ class CUDABackend(BaseBackend):
             opt_level = ['--opt-level', '0'] if os.environ.get("DISABLE_PTXAS_OPT", "0") == "1" else []
             ptxas_cmd = [ptxas, *line_info, *fmad, '-v', *opt_level, f'--gpu-name={arch}', fsrc.name, '-o', fbin]
             try:
-                subprocess.run(ptxas_cmd, check=True, close_fds=True, stdout=flog, stderr=flog)
+                # close_fds=True on Windows and False on Linux, see https://github.com/triton-lang/triton/pull/4357
+                # On Windows, both stdout and stderr need to be redirected to flog
+                subprocess.run(ptxas_cmd, check=True, close_fds=True if os.name == 'nt' else False, stdout=flog, stderr=flog)
             except subprocess.CalledProcessError as e:
                 with open(flog.name) as log_file:
                     log = log_file.read()

triton/backends/nvidia/driver.py

@@ -74,6 +74,10 @@ def compile_module_from_src(src, name):
             so = _build(name, src_path, tmpdir, library_dirs(), include_dir, libraries)
             with open(so, "rb") as f:
                 cache_path = cache.put(f.read(), f"{name}.{ext}", binary=True)
+
+    # Loading module with relative path may cause error
+    cache_path = os.path.abspath(cache_path)
+
     import importlib.util
     spec = importlib.util.spec_from_file_location(name, cache_path)
     mod = importlib.util.module_from_spec(spec)

triton/runtime/autotuner.py

@@ -36,7 +36,10 @@ class Autotuner(KernelInterface):
             'prune_num_stages_by'(optional): a function used to prune num_stages. It takes configs:List[Config] as its input, and returns pruned configs.
         """
         if not configs:
-            self.configs = [Config({}, num_warps=4, num_stages=3, num_ctas=1)]
+            self.configs = [
+                Config({}, num_warps=4, num_stages=3, num_ctas=1, num_buffers_warp_spec=0, num_consumer_groups=0,
+                       reg_dec_producer=0, reg_inc_consumer=0)
+            ]
         else:
             self.configs = configs
         self.keys = key
@@ -269,11 +272,16 @@ class Config:
                     function are args.
     """
 
-    def __init__(self, kwargs, num_warps=4, num_stages=3, num_ctas=1, maxnreg=None, pre_hook=None):
+    def __init__(self, kwargs, num_warps=4, num_stages=3, num_ctas=1, num_buffers_warp_spec=0, num_consumer_groups=0,
+                 reg_dec_producer=0, reg_inc_consumer=0, maxnreg=None, pre_hook=None):
         self.kwargs = kwargs
         self.num_warps = num_warps
         self.num_ctas = num_ctas
         self.num_stages = num_stages
+        self.num_buffers_warp_spec = num_buffers_warp_spec
+        self.num_consumer_groups = num_consumer_groups
+        self.reg_dec_producer = reg_dec_producer
+        self.reg_inc_consumer = reg_inc_consumer
         self.maxnreg = maxnreg
         self.pre_hook = pre_hook
 
@@ -285,6 +293,10 @@ class Config:
                     ("num_warps", self.num_warps),
                     ("num_ctas", self.num_ctas),
                     ("num_stages", self.num_stages),
+                    ("num_buffers_warp_spec", self.num_buffers_warp_spec),
+                    ("num_consumer_groups", self.num_consumer_groups),
+                    ("reg_dec_producer", self.reg_dec_producer),
+                    ("reg_inc_consumer", self.reg_inc_consumer),
                     ("maxnreg", self.maxnreg),
                 ) if v is not None
             }
@@ -297,6 +309,10 @@ class Config:
         res.append(f"num_warps: {self.num_warps}")
         res.append(f"num_ctas: {self.num_ctas}")
         res.append(f"num_stages: {self.num_stages}")
+        res.append(f"num_buffers_warp_spec: {self.num_buffers_warp_spec}")
+        res.append(f"num_consumer_groups: {self.num_consumer_groups}")
+        res.append(f"reg_dec_producer: {self.reg_dec_producer}")
+        res.append(f"reg_inc_consumer: {self.reg_inc_consumer}")
         res.append(f"maxnreg: {self.maxnreg}")
         return ", ".join(res)
 

triton/runtime/build.py

@@ -1,3 +1,4 @@
+import functools
 import sysconfig
 import os
 import shutil
@@ -7,8 +8,12 @@ if os.name == "nt":
     from triton.windows_utils import find_msvc_winsdk, find_python
 
 
+@functools.cache
 def get_cc():
     cc = os.environ.get("CC")
+    if cc is None:
+        # Find and check MSVC and Windows SDK from environment variables set by Launch-VsDevShell.ps1 or VsDevCmd.bat
+        cc, _, _ = find_msvc_winsdk(env_only=True)
     if cc is None:
         # Bundled TinyCC
         cc = os.path.join(sysconfig.get_paths()["platlib"], "triton", "runtime", "tcc", "tcc.exe")
@@ -75,9 +80,15 @@ def _build(name, src, srcdir, library_dirs, include_dirs, libraries):
     if "python3" not in libraries:
         libraries += ["python3"]
     if is_msvc(cc):
-        msvc_winsdk_inc_dirs, msvc_winsdk_lib_dirs = find_msvc_winsdk()
+        _, msvc_winsdk_inc_dirs, msvc_winsdk_lib_dirs = find_msvc_winsdk()
         include_dirs += msvc_winsdk_inc_dirs
         library_dirs += msvc_winsdk_lib_dirs
     cc_cmd = _cc_cmd(cc, src, so, include_dirs, library_dirs, libraries)
-    subprocess.check_call(cc_cmd, stdout=subprocess.DEVNULL)
+
+    try:
+        ret = subprocess.check_call(cc_cmd)
+    except Exception as e:
+        print("Failed to compile. cc_cmd:", cc_cmd)
+        raise e
+
     return so

triton/runtime/cache.py

@@ -131,7 +131,15 @@ class FileCacheManager(CacheManager):
             f.write(data)
         # Replace is guaranteed to be atomic on POSIX systems if it succeeds
         # so filepath cannot see a partial write
-        os.replace(temp_path, filepath)
+        try:
+            os.replace(temp_path, filepath)
+        except PermissionError:
+            # Ignore PermissionError on Windows because it happens when another process already
+            # put a file into the cache and locked it by opening it.
+            if os.name == "nt":
+                os.remove(temp_path)
+            else:
+                raise
         os.removedirs(temp_dir)
         return filepath
 

triton/tools/extra/cuda/compile.c

@@ -65,4 +65,5 @@ CUresult {kernel_name}(CUstream stream, {signature}) {{
     // TODO: shared memory
     if(gX * gY * gZ > 0)
       return cuLaunchKernel({kernel_name}_func, gX, gY, gZ, {num_warps} * 32, 1, 1, {shared}, stream, args, NULL);
+    return (CUresult)NULL;
 }}

triton/windows_utils.py

@@ -57,12 +57,31 @@ def check_msvc(msvc_base_path: Path, version: str) -> bool:
     return all(
         x.exists()
         for x in [
+            msvc_base_path / version / "bin" / "Hostx64" / "x64" / "cl.exe",
             msvc_base_path / version / "include" / "vcruntime.h",
             msvc_base_path / version / "lib" / "x64" / "vcruntime.lib",
         ]
     )
 
 
+def find_msvc_env() -> tuple[Optional[Path], Optional[str]]:
+    msvc_base_path = os.getenv("VCINSTALLDIR")
+    if msvc_base_path is None:
+        return None, None
+    msvc_base_path = Path(msvc_base_path) / "Tools" / "MSVC"
+
+    version = os.getenv("VCToolsVersion")
+    if not check_msvc(msvc_base_path, version):
+        warnings.warn(
+            f"Environment variables VCINSTALLDIR = {os.getenv('VCINSTALLDIR')}, "
+            f"VCToolsVersion = {os.getenv('VCToolsVersion')} are set, "
+            "but this MSVC installation is incomplete."
+        )
+        return None, None
+
+    return msvc_base_path, version
+
+
 def find_msvc_vswhere() -> tuple[Optional[Path], Optional[str]]:
     vswhere_path = find_in_program_files(
         r"Microsoft Visual Studio\Installer\vswhere.exe"
@@ -144,20 +163,28 @@ def find_msvc_hardcoded() -> tuple[Optional[Path], Optional[str]]:
     return None, None
 
 
-def find_msvc() -> tuple[list[str], list[str]]:
-    msvc_base_path, version = find_msvc_vswhere()
-    if msvc_base_path is None:
-        msvc_base_path, version = find_msvc_envpath()
-    if msvc_base_path is None:
-        msvc_base_path, version = find_msvc_hardcoded()
-    if msvc_base_path is None:
+def find_msvc(env_only: bool) -> tuple[Optional[str], list[str], list[str]]:
+    if env_only:
+        fs = [find_msvc_env]
+    else:
+        fs = [
+            find_msvc_env,
+            find_msvc_vswhere,
+            find_msvc_envpath,
+            find_msvc_hardcoded,
+        ]
+    for f in fs:
+        msvc_base_path, version = f()
+        if msvc_base_path:
+            return (
+                str(msvc_base_path / version / "bin" / "Hostx64" / "x64" / "cl.exe"),
+                [str(msvc_base_path / version / "include")],
+                [str(msvc_base_path / version / "lib" / "x64")],
+            )
+
+    if not env_only:
         warnings.warn("Failed to find MSVC.")
-        return [], []
-
-    return (
-        [str(msvc_base_path / version / "include")],
-        [str(msvc_base_path / version / "lib" / "x64")],
-    )
+    return None, [], []
 
 
 def check_winsdk(winsdk_base_path: Path, version: str) -> bool:
@@ -170,6 +197,26 @@ def check_winsdk(winsdk_base_path: Path, version: str) -> bool:
     )
 
 
+def find_winsdk_env() -> tuple[Optional[Path], Optional[str]]:
+    winsdk_base_path = os.getenv("WindowsSdkDir")
+    if winsdk_base_path is None:
+        return None, None
+    winsdk_base_path = Path(winsdk_base_path)
+
+    version = os.getenv("WindowsSDKVersion")
+    if version:
+        version = version.rstrip("\\")
+    if not check_winsdk(winsdk_base_path, version):
+        warnings.warn(
+            f"Environment variables WindowsSdkDir = {os.getenv('WindowsSdkDir')}, "
+            f"WindowsSDKVersion = {os.getenv('WindowsSDKVersion')} are set, "
+            "but this Windows SDK installation is incomplete."
+        )
+        return None, None
+
+    return winsdk_base_path, version
+
+
 def find_winsdk_registry() -> tuple[Optional[Path], Optional[str]]:
     try:
         reg = winreg.ConnectRegistry(None, winreg.HKEY_LOCAL_MACHINE)
@@ -212,32 +259,46 @@ def find_winsdk_hardcoded() -> tuple[Optional[Path], Optional[str]]:
     return winsdk_base_path, version
 
 
-def find_winsdk() -> tuple[list[str], list[str]]:
-    winsdk_base_path, version = find_winsdk_registry()
-    if winsdk_base_path is None:
-        winsdk_base_path, version = find_winsdk_hardcoded()
-    if winsdk_base_path is None:
+def find_winsdk(env_only: bool) -> tuple[list[str], list[str]]:
+    if env_only:
+        fs = [find_winsdk_env]
+    else:
+        fs = [
+            find_winsdk_env,
+            find_winsdk_registry,
+            find_winsdk_hardcoded,
+        ]
+    for f in fs:
+        winsdk_base_path, version = f()
+        if winsdk_base_path:
+            return (
+                [
+                    str(winsdk_base_path / "Include" / version / "shared"),
+                    str(winsdk_base_path / "Include" / version / "ucrt"),
+                    str(winsdk_base_path / "Include" / version / "um"),
+                ],
+                [
+                    str(winsdk_base_path / "Lib" / version / "ucrt" / "x64"),
+                    str(winsdk_base_path / "Lib" / version / "um" / "x64"),
+                ],
+            )
+
+    if not env_only:
         warnings.warn("Failed to find Windows SDK.")
-        return [], []
-
-    return (
-        [
-            str(winsdk_base_path / "Include" / version / "shared"),
-            str(winsdk_base_path / "Include" / version / "ucrt"),
-            str(winsdk_base_path / "Include" / version / "um"),
-        ],
-        [
-            str(winsdk_base_path / "Lib" / version / "ucrt" / "x64"),
-            str(winsdk_base_path / "Lib" / version / "um" / "x64"),
-        ],
-    )
+    return [], []
 
 
 @functools.cache
-def find_msvc_winsdk() -> tuple[list[str], list[str]]:
-    msvc_inc_dirs, msvc_lib_dirs = find_msvc()
-    winsdk_inc_dirs, winsdk_lib_dirs = find_winsdk()
-    return msvc_inc_dirs + winsdk_inc_dirs, msvc_lib_dirs + winsdk_lib_dirs
+def find_msvc_winsdk(
+    env_only: bool = False,
+) -> tuple[Optional[str], list[str], list[str]]:
+    msvc_bin_path, msvc_inc_dirs, msvc_lib_dirs = find_msvc(env_only)
+    winsdk_inc_dirs, winsdk_lib_dirs = find_winsdk(env_only)
+    return (
+        msvc_bin_path,
+        msvc_inc_dirs + winsdk_inc_dirs,
+        msvc_lib_dirs + winsdk_lib_dirs,
+    )
 
 
 @functools.cache

{triton_windows-3.3.0a0.post16.dist-info → triton_windows-3.3.0a0.post18.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: triton-windows
-Version: 3.3.0a0.post16
+Version: 3.3.0a0.post18
 Summary: A language and compiler for custom Deep Learning operations
 Home-page: https://github.com/woct0rdho/triton-windows
 Author: Philippe Tillet, Dian Wu

{triton_windows-3.3.0a0.post16.dist-info → triton_windows-3.3.0a0.post18.dist-info}/RECORD

@@ -3,12 +3,12 @@ triton/_internal_testing.py,sha256=OBY28huiEWItqGgiukgZzHLLaSbS8yj9kdhn_u562Yg,5
 triton/_utils.py,sha256=5RiCLwW14w0Q3mdZ-9yz-VO5KiSexNj9xeDt4gaNsvE,1014
 triton/errors.py,sha256=8WfnuRKLG578mgY6cBA3ECruVMf9ULEKFNgRcJ6IhWM,89
 triton/testing.py,sha256=ivFf1Fq9frmfVahaVUp0bgJxmvVZNACZfj3Sai6zfAs,20048
-triton/windows_utils.py,sha256=aQMItmuZNXaki8zSB7HMvhy1RsXlmz1GcO-lORm1IIk,10852
-triton/_C/libtriton.pyd,sha256=ZqauD5-vXP3XsuVmqNJecoJX1crAkDGRAwThWddn1QY,86825472
+triton/windows_utils.py,sha256=YUl-1QbLINQRaAAMNYPjLiTFZlVsCNi9mTFck5aemwk,12778
+triton/_C/libtriton.pyd,sha256=CHlRXTuChUtksf319rUsjMpt_uDJws61neyTqpBb2cU,87276032
 triton/backends/__init__.py,sha256=opAo_vgEMt3tLO_bYFrYGksnIu0qohbmyuu_s3-rNAs,1595
 triton/backends/compiler.py,sha256=ymaG0kpveAuESbQ9QZ0RyXjr0Aq4el_G5XGYogJ2gNA,3588
 triton/backends/driver.py,sha256=AN60upJlPgia0JwvZ8vIVgLITNPuI0fdz8zMIIHPpF4,1450
-triton/backends/amd/compiler.py,sha256=7Ej6QPi85Jwdx7E27d7A--a00OipDkpgiL3JNnCwoTo,19007
+triton/backends/amd/compiler.py,sha256=27jurEV7tH5J6BHtOXrdPJyCMYIPHo5G4Op_O66gv4E,19135
 triton/backends/amd/driver.c,sha256=obiiiPndny5NyhUcJ8iyrVHrXU1ruLpLGd_LgaKQEbU,8459
 triton/backends/amd/driver.py,sha256=p8FcTiAq-829p2gRQZ5sPE1_d1SQQ_sOhb5WxI6rXME,20379
 triton/backends/amd/include/hip/channel_descriptor.h,sha256=gTYe7SzIg-m3ThOQY2vr5Rh6-uWvUP_d37v8F4T2Q14,1773
@@ -112,9 +112,9 @@ triton/backends/amd/lib/asanrtl.bc,sha256=1xv2RlU3WvbdsghHlmhwiHewGM2B5dKts5bERM
 triton/backends/amd/lib/ockl.bc,sha256=wQKCzkKukIHbu0lyjKUYlhndc7S27xto6L54J0Bn-C0,246124
 triton/backends/amd/lib/ocml.bc,sha256=UPNTXW0gCXUNB-c6orSYwb-mz9_mjUc7zny_vfFza44,205964
 triton/backends/nvidia/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-triton/backends/nvidia/compiler.py,sha256=iHbIl-HXSgD3S5eJIXbQhr-Cyk9Gq_g8RXDFZxAGUF4,18444
+triton/backends/nvidia/compiler.py,sha256=4BnLUapsoh8lsQPYlm-_ZCyKvU3DgJY3NgDMo8leV9c,19996
 triton/backends/nvidia/driver.c,sha256=RWzdn9_RWaaVxXKtdrvxoRg4yR7WSH7ELRyjGDYKwBM,18880
-triton/backends/nvidia/driver.py,sha256=BcEzr0_ONM8H_DKoCbNuO1jRpBM_XHKBgt_Y2bG9aZw,20902
+triton/backends/nvidia/driver.py,sha256=cbNUIsppHawGsqf3ctuqc0nRM_kbsMCsH1osWv0fKro,21005
 triton/backends/nvidia/bin/ptxas.exe,sha256=iCva9hIYg-Q2NybchwaQJFkwDzNiliFOTDdZrHPLb6A,24732160
 triton/backends/nvidia/include/cuda.h,sha256=Fn44OjeRImxegJ39apYUspseEfTWNGwpqSGUOnHj5WY,1183268
 triton/backends/nvidia/lib/libdevice.10.bc,sha256=XC-uN8huaMOjhgWpX1EtfRLV89uYYxC-R_VzBKpype4,473728
@@ -141,9 +141,9 @@ triton/language/extra/hip/__init__.py,sha256=ieSER4LeX9_0horChGUUVwpuKAprkuka8uG
 triton/language/extra/hip/libdevice.py,sha256=EVraUfeXzQmN3F5Lleg2mohVcbFWOWlLaAH1nkbqtV4,16841
 triton/runtime/__init__.py,sha256=mKL5cqIBDUw2WO80NRCh4s1G8KYaqgM59TTAbTkPPjQ,621
 triton/runtime/_allocation.py,sha256=zaW4B7I7c-2rkVuN7IZaUB6IQSI1t4FvnTPZH-r7DTk,798
-triton/runtime/autotuner.py,sha256=h1I6TDPPYtMvuwOkeprcSD3VKKGXk952jXN64oPIFXs,16844
-triton/runtime/build.py,sha256=6Ie4gn4-Zk1DIW3o-Nd002IhtEm_yikXh9o6tUm6iao,3279
-triton/runtime/cache.py,sha256=Vs9l4DeRhKXyB-hfDBrp5v85OO6vkkXTHTVjZ7zu0xg,10260
+triton/runtime/autotuner.py,sha256=0ku0wjPo8xOvom6P4uEVZHsgPkxAFOqP1LjUVue0HLM,17854
+triton/runtime/build.py,sha256=1kRGMsDDg7NHdxKEPmNJ3SQaa_UnBckJF5KfFOwoBbI,3593
+triton/runtime/cache.py,sha256=uoU1UH3HPxkxT0r-69HIZgLBls3T-TDxJedRULX9lbM,10583
 triton/runtime/driver.py,sha256=VZ-883Xri71R72lHB6usIpLo3gGLbZJkAlLP3ewWSpc,1509
 triton/runtime/errors.py,sha256=CwfJXciwel_-K3BfQfKUpLPDWrSyTnGsfJkqJojrdfQ,1052
 triton/runtime/interpreter.py,sha256=RVzNGSXyLlg14LgjJ5BZ3qbqfK49FccG-pXaIXXXE2g,59331
@@ -245,9 +245,9 @@ triton/tools/disasm.py,sha256=BBO4bALdLcWgWDLhQdYHLlTx3oo8g_d8maeE_Uu-FmU,5088
 triton/tools/experimental_descriptor.py,sha256=0Wqy96Cc6YLh9o0eTknW-Lfvha6lfRSfe8bswkcPHMs,1260
 triton/tools/link.py,sha256=u7qtfZRLriZkAMEGNvj8YF-k1cthmLL7BwHYqBgT63E,11871
 triton/tools/mxfp.py,sha256=YQdpBrGkOVNOtnLeRjMCeVFHWkSwUubGeWsItIjO8TU,11737
-triton/tools/extra/cuda/compile.c,sha256=Me7beHPc6WNTwjg85H84DUMCRu4KJdVK2hNNgvlhBZ4,2126
+triton/tools/extra/cuda/compile.c,sha256=TdIENsqk6wrvv1C4Mk-sq9keXe3SJuMQcf0UpxmjNZk,2153
 triton/tools/extra/cuda/compile.h,sha256=n9QKIFZTL4RSsiXtAxBP9XGSnxjyaevQQ9bBpwDsvAg,332
-triton_windows-3.3.0a0.post16.dist-info/METADATA,sha256=soxppI1GtZ5SPZoHEK6khzoUWAnqHdzSINwOc5AsKLs,1629
-triton_windows-3.3.0a0.post16.dist-info/WHEEL,sha256=kynURWPI8D6MJZUv3wOGtm1TQb0_WLqkF127RFCfG9M,101
-triton_windows-3.3.0a0.post16.dist-info/top_level.txt,sha256=KhMzHYsArnZ3IkjAQ-xLnx1n_FjvEpJNelg2xPiDl-U,254
-triton_windows-3.3.0a0.post16.dist-info/RECORD,,
+triton_windows-3.3.0a0.post18.dist-info/METADATA,sha256=2h8i8geeJlfffh1CwstG76Xf4jcC2yuGasvloSjd43I,1629
+triton_windows-3.3.0a0.post18.dist-info/WHEEL,sha256=_ZWIY2n7n6SpiuIFl1-RvcMp4Ty36T57FKf-7NzqZHM,101
+triton_windows-3.3.0a0.post18.dist-info/top_level.txt,sha256=KhMzHYsArnZ3IkjAQ-xLnx1n_FjvEpJNelg2xPiDl-U,254
+triton_windows-3.3.0a0.post18.dist-info/RECORD,,

{triton_windows-3.3.0a0.post16.dist-info → triton_windows-3.3.0a0.post18.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (77.0.1)
+Generator: setuptools (78.1.0)
 Root-Is-Purelib: false
 Tag: cp311-cp311-win_amd64