triton-windows 3.3.1.post19__cp312-cp312-win_amd64.whl → 3.5.0.post21__cp312-cp312-win_amd64.whl

triton/backends/nvidia/driver.py

@@ -1,37 +1,37 @@
 import functools
+import operator
 import os
-import sysconfig
-import hashlib
 import subprocess
-import tempfile
+import triton
+import re
 from pathlib import Path
-from triton.runtime.build import _build
-from triton.runtime.cache import get_cache_manager
+from triton import knobs
+from triton.runtime.build import compile_module_from_src
 from triton.runtime import _allocation
 from triton.backends.compiler import GPUTarget
 from triton.backends.driver import GPUDriver
 
 dirname = os.path.dirname(os.path.realpath(__file__))
-include_dir = [os.path.join(dirname, "include")]
+include_dirs = [os.path.join(dirname, "include")]
 if os.name == "nt":
     from triton.windows_utils import find_cuda
     _, cuda_inc_dirs, _ = find_cuda()
-    include_dir += cuda_inc_dirs
+    include_dirs += cuda_inc_dirs
 libdevice_dir = os.path.join(dirname, "lib")
 libraries = ['cuda']
+PyCUtensorMap = None
 
 
 @functools.lru_cache()
 def libcuda_dirs():
-    env_libcuda_path = os.getenv("TRITON_LIBCUDA_PATH")
-    if env_libcuda_path:
+    if env_libcuda_path := knobs.nvidia.libcuda_path:
         return [env_libcuda_path]
 
     if os.name == "nt":
         _, _, cuda_lib_dirs = find_cuda()
         return cuda_lib_dirs
 
-    libs = subprocess.check_output(["/sbin/ldconfig", "-p"]).decode()
+    libs = subprocess.check_output(["/sbin/ldconfig", "-p"]).decode(errors="ignore")
     # each line looks like the following:
     # libcuda.so.1 (libc6,x86-64) => /lib/x86_64-linux-gnu/libcuda.so.1
     locs = [line.split()[-1] for line in libs.splitlines() if "libcuda.so.1" in line]
@@ -55,36 +55,6 @@ def library_dirs():
     return [libdevice_dir, *libcuda_dirs()]
 
 
-@functools.lru_cache()
-def platform_key():
-    from platform import machine, system, architecture
-    return ",".join([machine(), system(), *architecture()])
-
-
-def compile_module_from_src(src, name):
-    key = hashlib.sha256((src + platform_key()).encode("utf-8")).hexdigest()
-    cache = get_cache_manager(key)
-    ext = sysconfig.get_config_var("EXT_SUFFIX").split(".")[-1]
-    cache_path = cache.get_file(f"{name}.{ext}")
-    if cache_path is None:
-        with tempfile.TemporaryDirectory() as tmpdir:
-            src_path = os.path.join(tmpdir, f"{name}.c")
-            with open(src_path, "w") as f:
-                f.write(src)
-            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)
-    spec.loader.exec_module(mod)
-    return mod
-
-
 # ------------------------
 # Utils
 # ------------------------
@@ -98,13 +68,20 @@ class CudaUtils(object):
         return cls.instance
 
     def __init__(self):
-        mod = compile_module_from_src(Path(os.path.join(dirname, "driver.c")).read_text(), "cuda_utils")
+        mod = compile_module_from_src(
+            src=Path(os.path.join(dirname, "driver.c")).read_text(),
+            name="cuda_utils",
+            library_dirs=library_dirs(),
+            include_dirs=include_dirs,
+            libraries=libraries,
+        )
+        global PyCUtensorMap
+        PyCUtensorMap = mod.PyCUtensorMap
         self.load_binary = mod.load_binary
         self.get_device_properties = mod.get_device_properties
         self.cuOccupancyMaxActiveClusters = mod.cuOccupancyMaxActiveClusters
         self.set_printf_fifo_size = mod.set_printf_fifo_size
-        self.fill_1d_tma_descriptor = mod.fill_1d_tma_descriptor
-        self.fill_2d_tma_descriptor = mod.fill_2d_tma_descriptor
+        self.fill_tma_descriptor = mod.fill_tma_descriptor
 
 
 # ------------------------
@@ -115,32 +92,95 @@ class CudaUtils(object):
 def ty_to_cpp(ty):
     if ty[0] == '*':
         return "CUdeviceptr"
+    if ty.startswith("tensordesc"):
+        return "CUtensorMap"
     return {
-        "i1": "int32_t",
+        "i1": "int8_t",
         "i8": "int8_t",
         "i16": "int16_t",
         "i32": "int32_t",
         "i64": "int64_t",
-        "u1": "uint32_t",
+        "u1": "uint8_t",
         "u8": "uint8_t",
         "u16": "uint16_t",
         "u32": "uint32_t",
         "u64": "uint64_t",
-        "fp16": "float",
-        "bf16": "float",
-        "fp32": "float",
-        "f32": "float",
+        "fp16": "double",
+        "bf16": "double",
+        "fp32": "double",
+        "f32": "double",
         "fp64": "double",
         "nvTmaDesc": "CUtensorMap",
     }[ty]
 
 
-def make_launcher(constants, signature):
-
-    def _serialize_signature(sig):
+FLOAT_STORAGE_TYPE = {
+    "fp16": "uint16_t",
+    "bf16": "uint16_t",
+    "fp32": "uint32_t",
+    "f32": "uint32_t",
+    "fp64": "uint64_t",
+}
+FLOAT_PACK_FUNCTION = {
+    "fp16": "pack_fp16",
+    "bf16": "pack_bf16",
+    "fp32": "pack_fp32",
+    "f32": "pack_fp32",
+    "fp64": "pack_fp64",
+}
+
+_BASE_ARGS_FORMAT = "iiiKKppOOOOOO"
+_BASE_ARGS_FORMAT_LEN = len(_BASE_ARGS_FORMAT)
+
+
+def make_launcher(constants, signature, tensordesc_meta):
+
+    def _expand_signature(signature):
+        output = []
+        tensordesc_idx = 0
+        # Expand tensor descriptor arguments into either nvTmaDesc, shape and
+        # strides, or base pointer, shape and strides depending on whether the
+        # kernel was lowered to use the nvTmaDesc or not.
+        for sig in signature:
+            if isinstance(sig, str) and sig.startswith("tensordesc"):
+                meta = tensordesc_meta[tensordesc_idx] if tensordesc_meta else None
+                tensordesc_idx += 1
+
+                match = re.match("tensordesc<([^[>]*)\\[([^]]*)\\]", sig)
+                dtype = match.group(1)
+                shape = match.group(2)
+                ndim = shape.count(",") + 1
+
+                if meta is None:
+                    output.append("*" + dtype)
+                    # Currently the host side tensor descriptors get passed in as a
+                    # tensor desc, shape, and strides. We have no way to use these
+                    # shape and strides when processing tensor descriptors which is
+                    # why we provide our own decomposition above. Sadly this means
+                    # we have to pass the shape and strides twice.
+                    for _ in range(2 * ndim):
+                        output.append("i64")
+                    output.append("i1")
+                else:
+                    output.append("nvTmaDesc")
+
+                for _ in range(ndim):
+                    output.append("i32")
+                for _ in range(ndim):
+                    output.append("i64")
+            else:
+                output.append(sig)
+
+        assert not tensordesc_meta or tensordesc_idx == len(tensordesc_meta)
+        return output
+
+    def _flatten_signature(sig, output):
+        # Flatten tuples
         if isinstance(sig, tuple):
-            return ','.join(map(_serialize_signature, sig))
-        return sig
+            for x in sig:
+                _flatten_signature(x, output)
+        else:
+            output.append(sig)
 
     def _extracted_type(ty):
         if isinstance(ty, tuple):
@@ -160,8 +200,9 @@ def make_launcher(constants, signature):
             return "O"
         if ty in ("constexpr", "nvTmaDesc"):
             return "O"
+        if ty.startswith("tensordesc"):
+            return "O"
         return {
-            "float": "f",
             "double": "d",
             "long": "l",
             "int8_t": "b",
@@ -174,19 +215,34 @@ def make_launcher(constants, signature):
             "uint64_t": "K",
         }[ty_to_cpp(ty)]
 
+    expand_signature = _expand_signature(signature.values())
+    signature = {i: s for i, s in enumerate(expand_signature)}
+
     args_format = ''.join([format_of(ty) for ty in signature.values()])
-    format = "iiiKKpOOOOO" + args_format
-    signature = ','.join(map(_serialize_signature, signature.values()))
-    signature = list(filter(bool, signature.split(',')))
-    signature = {i: s for i, s in enumerate(signature)}
+    format = _BASE_ARGS_FORMAT + args_format
+
+    flat_signature = []
+    for sig in signature.values():
+        _flatten_signature(sig, flat_signature)
+    signature = {i: s for i, s in enumerate(flat_signature)}
     args_list = ', ' + ', '.join(f"&_arg{i}" for i, ty in signature.items()) if len(signature) > 0 else ''
     # Record the end of regular arguments;
     # subsequent arguments are architecture-specific descriptors, such as tensor descriptors for CUDA.
-    arg_decls = ', '.join(f"{ty_to_cpp(ty)} arg{i}" for i, ty in signature.items() if ty != "constexpr")
+    arg_decl_list = []
+    for i, ty in signature.items():
+        if ty == "constexpr":
+            continue
+        if ty in FLOAT_STORAGE_TYPE:
+            arg_decl_list.append(f"{FLOAT_STORAGE_TYPE[ty]} arg{i}")
+        else:
+            arg_decl_list.append(f"{ty_to_cpp(ty)} arg{i}")
+    arg_decls = ', '.join(arg_decl_list)
     internal_args_list = []
     for i, ty in signature.items():
         if ty[0] == "*":
             internal_args_list.append(f"ptr_info{i}.dev_ptr")
+        elif ty in FLOAT_STORAGE_TYPE:
+            internal_args_list.append(f"_arg{i}_storage")
         elif ty == "nvTmaDesc":
             # Note: we have to dereference the pointer
             internal_args_list.append(f"*tma_ptr{i}")
@@ -205,15 +261,17 @@ def make_launcher(constants, signature):
         f"CUtensorMap* tma_ptr{i} = getTmaDesc(_arg{i}); if (!tma_ptr{i}) return NULL;" for i, ty in signature.items()
         if ty == "nvTmaDesc"
     ]
+    float_storage_decls = [
+        f"{FLOAT_STORAGE_TYPE[ty]} _arg{i}_storage = {FLOAT_PACK_FUNCTION[ty]}(_arg{i});"
+        for i, ty in signature.items()
+        if ty in FLOAT_STORAGE_TYPE
+    ]
     params = [f"&arg{i}" for i, ty in signature.items() if ty != "constexpr"]
     params.append("&global_scratch")
+    params.append("&profile_scratch")
     src = f"""
 #define _CRT_SECURE_NO_WARNINGS
 #include \"cuda.h\"
-#include <stdbool.h>
-#define PY_SSIZE_T_CLEAN
-#define Py_LIMITED_API 0x03090000
-#include <Python.h>
 
 #ifndef _WIN32
 #include <dlfcn.h>
@@ -222,6 +280,16 @@ def make_launcher(constants, signature):
 #include <windows.h>
 #endif
 
+#include <stdbool.h>
+#include <stdlib.h>
+#define PY_SSIZE_T_CLEAN
+#include <Python.h>
+
+typedef struct {{
+  PyObject_HEAD
+  _Alignas(128) CUtensorMap tensorMap;
+}} PyCUtensorMapObject;
+
 static inline void gpuAssert(CUresult code, const char *file, int line)
 {{
    if (code != CUDA_SUCCESS)
@@ -282,67 +350,65 @@ static cuLaunchKernelEx_t getLaunchKernelExHandle() {{
 }}
 #endif
 
-static void _launch(int gridX, int gridY, int gridZ, int num_warps, int num_ctas, int launch_cooperative_grid, int clusterDimX, int clusterDimY, int clusterDimZ, int shared_memory, CUstream stream, CUfunction function, CUdeviceptr global_scratch{', ' + arg_decls if len(arg_decls) > 0 else ''}) {{
+static void _launch(int gridX, int gridY, int gridZ, int num_warps, int num_ctas, int launch_cooperative_grid, int launch_pdl, int clusterDimX, int clusterDimY, int clusterDimZ, int shared_memory, CUstream stream, CUfunction function, CUdeviceptr global_scratch, CUdeviceptr profile_scratch{', ' + arg_decls if len(arg_decls) > 0 else ''}) {{
   void *params[] = {{ {', '.join(params)} }};
   if (gridX*gridY*gridZ > 0) {{
-    if ((num_ctas == 1) && (0 == launch_cooperative_grid)) {{
-      CUDA_CHECK(cuLaunchKernel(function, gridX, gridY, gridZ, 32*num_warps, 1, 1, shared_memory, stream, params, 0));
-    }} else if ((num_ctas == 1) && (0 != launch_cooperative_grid)) {{
-      CUlaunchAttribute launchAttr[1];
+    // 4 attributes that we can currently pass maximum
+    CUlaunchAttribute launchAttr[4];
+    static cuLaunchKernelEx_t cuLaunchKernelExHandle = NULL;
+    if (cuLaunchKernelExHandle == NULL) {{
+      cuLaunchKernelExHandle = getLaunchKernelExHandle();
+    }}
+    CUlaunchConfig config;
+    config.gridDimX = gridX;
+    config.gridDimY = gridY;
+    config.gridDimZ = gridZ;
+
+    if (num_ctas != 1) {{
+      config.gridDimX *= clusterDimX;
+      config.gridDimY *= clusterDimY;
+      config.gridDimZ *= clusterDimZ;
+    }}
+
+    config.blockDimX = 32 * num_warps;
+    config.blockDimY = 1;
+    config.blockDimZ = 1;
+    config.sharedMemBytes = shared_memory;
+    config.hStream = stream;
+    config.attrs = launchAttr;
+    int num_attrs = 0;
+
+    if (launch_pdl != 0) {{
+      CUlaunchAttribute pdlAttr = {{ .id = CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_STREAM_SERIALIZATION, .value = 1}};
+      launchAttr[num_attrs] = pdlAttr;
+      ++num_attrs;
+    }}
+
+    if (launch_cooperative_grid != 0) {{
       CUlaunchAttribute coopAttr = {{ .id = CU_LAUNCH_ATTRIBUTE_COOPERATIVE, .value = 1}};
-      launchAttr[0] = coopAttr;
-
-      CUlaunchConfig config;
-      config.gridDimX = gridX;
-      config.gridDimY = gridY;
-      config.gridDimZ = gridZ;
-      config.blockDimX = 32 * num_warps;
-      config.blockDimY = 1;
-      config.blockDimZ = 1;
-      config.sharedMemBytes = shared_memory;
-      config.hStream = stream;
-      config.attrs = launchAttr;
-      config.numAttrs = 1;
-
-      static cuLaunchKernelEx_t cuLaunchKernelExHandle = NULL;
-      if (cuLaunchKernelExHandle == NULL) {{
-        cuLaunchKernelExHandle = getLaunchKernelExHandle();
-      }}
-      CUDA_CHECK(cuLaunchKernelExHandle(&config, function, params, 0));
-
-    }} else {{
-      CUlaunchAttribute launchAttr[3];
-      launchAttr[0].id = CU_LAUNCH_ATTRIBUTE_CLUSTER_DIMENSION;
-      launchAttr[0].value.clusterDim.x = clusterDimX;
-      launchAttr[0].value.clusterDim.y = clusterDimY;
-      launchAttr[0].value.clusterDim.z = clusterDimZ;
-      launchAttr[1].id = CU_LAUNCH_ATTRIBUTE_CLUSTER_SCHEDULING_POLICY_PREFERENCE;
-      launchAttr[1].value.clusterSchedulingPolicyPreference = CU_CLUSTER_SCHEDULING_POLICY_SPREAD;
-
-      unsigned numAttrs = 2;
-      if (0 != launch_cooperative_grid) {{
-        CUlaunchAttribute coopAttr = {{ .id = CU_LAUNCH_ATTRIBUTE_COOPERATIVE, .value = 1}};
-        launchAttr[2] = coopAttr;
-        numAttrs = 3;
-      }}
-
-      CUlaunchConfig config;
-      config.gridDimX = gridX * clusterDimX;
-      config.gridDimY = gridY * clusterDimY;
-      config.gridDimZ = gridZ * clusterDimZ;
-      config.blockDimX = 32 * num_warps;
-      config.blockDimY = 1;
-      config.blockDimZ = 1;
-      config.sharedMemBytes = shared_memory;
-      config.hStream = stream;
-      config.attrs = launchAttr;
-      config.numAttrs = numAttrs;
-      static cuLaunchKernelEx_t cuLaunchKernelExHandle = NULL;
-      if (cuLaunchKernelExHandle == NULL) {{
-        cuLaunchKernelExHandle = getLaunchKernelExHandle();
-      }}
-      CUDA_CHECK(cuLaunchKernelExHandle(&config, function, params, 0));
+      launchAttr[num_attrs] = coopAttr;
+      ++num_attrs;
     }}
+
+    if (num_ctas != 1) {{
+      CUlaunchAttribute clusterAttr = {{}};
+      clusterAttr.id = CU_LAUNCH_ATTRIBUTE_CLUSTER_DIMENSION;
+      clusterAttr.value.clusterDim.x = clusterDimX;
+      clusterAttr.value.clusterDim.y = clusterDimY;
+      clusterAttr.value.clusterDim.z = clusterDimZ;
+      launchAttr[num_attrs] = clusterAttr;
+      ++num_attrs;
+
+      CUlaunchAttribute clusterSchedulingAttr = {{}};
+      clusterSchedulingAttr.id = CU_LAUNCH_ATTRIBUTE_CLUSTER_SCHEDULING_POLICY_PREFERENCE;
+      clusterSchedulingAttr.value.clusterSchedulingPolicyPreference = CU_CLUSTER_SCHEDULING_POLICY_SPREAD;
+      launchAttr[num_attrs] = clusterSchedulingAttr;
+      ++num_attrs;
+    }}
+
+    config.numAttrs = num_attrs;
+
+    CUDA_CHECK(cuLaunchKernelExHandle(&config, function, params, 0));
   }}
 }}
 
@@ -351,6 +417,9 @@ typedef struct _DevicePtrInfo {{
     bool valid;
 }} DevicePtrInfo;
 
+static PyObject* data_ptr_str = NULL;
+static PyObject* py_tensor_map_type = NULL;
+
 static inline DevicePtrInfo getPointer(PyObject *obj, int idx) {{
   DevicePtrInfo ptr_info;
   ptr_info.dev_ptr = 0;
@@ -363,37 +432,35 @@ static inline DevicePtrInfo getPointer(PyObject *obj, int idx) {{
     // valid nullptr
     return ptr_info;
   }}
-  PyObject *ptr = PyObject_GetAttrString(obj, "data_ptr");
-  if(ptr){{
-    PyObject *empty_tuple = PyTuple_New(0);
-    PyObject *ret = PyObject_Call(ptr, empty_tuple, NULL);
-    Py_DECREF(empty_tuple);
-    Py_DECREF(ptr);
-    if (!PyLong_Check(ret)) {{
-      PyErr_SetString(PyExc_TypeError, "data_ptr method of Pointer object must return 64-bit int");
-      ptr_info.valid = false;
-      return ptr_info;
-    }}
-    ptr_info.dev_ptr = PyLong_AsUnsignedLongLong(ret);
-    if(!ptr_info.dev_ptr)
-      return ptr_info;
-    uint64_t dev_ptr;
-    int status = cuPointerGetAttribute(&dev_ptr, CU_POINTER_ATTRIBUTE_DEVICE_POINTER, ptr_info.dev_ptr);
-    if (status == CUDA_ERROR_INVALID_VALUE) {{
-        PyErr_Format(PyExc_ValueError,
-                     "Pointer argument (at %d) cannot be accessed from Triton (cpu tensor?)", idx);
-        ptr_info.valid = false;
-    }} else if (status != CUDA_SUCCESS) {{
-        CUDA_CHECK(status);  // Catch any other cuda API errors
-        ptr_info.valid = false;
-    }}
-    ptr_info.dev_ptr = dev_ptr;
-    Py_DECREF(ret);  // Thanks ChatGPT!
+  PyObject *ret = PyObject_CallMethodNoArgs(obj, data_ptr_str);
+  if (!ret) {{
+    PyErr_SetString(PyExc_TypeError, "Pointer argument must be either uint64 or have data_ptr method");
+    ptr_info.valid = false;
+    goto cleanup;
+  }}
+  if (!PyLong_Check(ret)) {{
+    PyErr_SetString(PyExc_TypeError, "data_ptr method of Pointer object must return 64-bit int");
+    ptr_info.valid = false;
+    goto cleanup;
+  }}
+  ptr_info.dev_ptr = PyLong_AsUnsignedLongLong(ret);
+  if(!ptr_info.dev_ptr)
     return ptr_info;
+  uint64_t dev_ptr;
+  int status = cuPointerGetAttribute(&dev_ptr, CU_POINTER_ATTRIBUTE_DEVICE_POINTER, ptr_info.dev_ptr);
+  if (status == CUDA_ERROR_INVALID_VALUE) {{
+      PyErr_Format(PyExc_ValueError,
+                   "Pointer argument (at %d) cannot be accessed from Triton (cpu tensor?)", idx);
+      ptr_info.valid = false;
+  }} else if (status != CUDA_SUCCESS) {{
+      CUDA_CHECK(status);  // Catch any other cuda API errors
+      ptr_info.valid = false;
   }}
-  PyErr_SetString(PyExc_TypeError, "Pointer argument must be either uint64 or have data_ptr method");
-  ptr_info.valid = false;
+  ptr_info.dev_ptr = dev_ptr;
+cleanup:
+  Py_XDECREF(ret);
   return ptr_info;
+
 }}
 
 static inline CUtensorMap* getTmaDesc(PyObject *obj) {{
@@ -402,44 +469,18 @@ static inline CUtensorMap* getTmaDesc(PyObject *obj) {{
     return NULL;
   }}
 
-  PyObject *method_handle = PyObject_GetAttrString(obj, "tma_desc_cpu_ptr");
-  if (!method_handle) {{
-    PyErr_SetString(PyExc_TypeError, "tma_desc_cpu_ptr() method does not exist");
-    return NULL;
-  }}
-
-  PyObject *empty_tuple = PyTuple_New(0);
-  if (!empty_tuple) {{
-    Py_DECREF(method_handle);
-    PyErr_SetString(PyExc_SystemError, "Internal Python error!");
-    return NULL;
-  }}
-  PyObject *method_ret = PyObject_Call(method_handle, empty_tuple, NULL);
-  Py_DECREF(empty_tuple);
-  Py_DECREF(method_handle);
-  if (!method_ret) {{
-    PyErr_SetString(PyExc_SystemError, "Internal Python error!");
-    return NULL;
-  }}
-
-  if (!PyLong_Check(method_ret)) {{
-    PyErr_SetString(PyExc_TypeError, "tma_desc_cpu_ptr() must return 64-bit int");
-    Py_DECREF(method_ret);
+if (Py_TYPE(obj) != (PyTypeObject*)py_tensor_map_type) {{
+    PyErr_Format(PyExc_TypeError, "object must be of type PyCUtensorMap, got %s", Py_TYPE(obj)->tp_name);
     return NULL;
-  }}
+}}
 
-  uint64_t ptr_as_uint = PyLong_AsUnsignedLongLong(method_ret);
-  Py_DECREF(method_ret);
-  if (!ptr_as_uint) {{
-    PyErr_SetString(PyExc_ValueError, "received NULL ptr from tma_desc_cpu_ptr()");
-    return NULL;
-  }}
-  if (ptr_as_uint % 64 != 0) {{
-    PyErr_SetString(PyExc_ValueError, "tma_desc_cpu_ptr() must be 64-byte aligned");
+  CUtensorMap* map = &((PyCUtensorMapObject*)obj)->tensorMap;
+  uintptr_t align_128 = (uintptr_t)map & (128 - 1);
+  if (align_128 != 0) {{
+    PyErr_Format(PyExc_ValueError, "CUtensorMap must be aligned to 128B, but got (&map) mod 128 = %ld", align_128);
     return NULL;
   }}
-
-  return (CUtensorMap*)(ptr_as_uint);
+  return map;
 }}
 
 static void ensureCudaContext() {{
@@ -454,6 +495,32 @@ static void ensureCudaContext() {{
   }}
 }}
 
+static uint16_t pack_fp16(double f) {{
+    uint16_t result;
+    // from https://github.com/python/pythoncapi-compat
+#if 0x030600B1 <= PY_VERSION_HEX && PY_VERSION_HEX <= 0x030B00A1 && !defined(PYPY_VERSION)
+    _PyFloat_Pack2(f, (unsigned char*)&result, 1);
+#else
+    PyFloat_Pack2(f, (unsigned char*)&result, 1);
+#endif
+    return result;
+}}
+
+static uint16_t pack_bf16(double f) {{
+    float f32 = (float)f;
+    uint32_t u32 = *(uint32_t*)&f32;
+    return (uint16_t)(u32 >> 16);
+}}
+
+static uint32_t pack_fp32(double f) {{
+    float f32 = (float)f;
+    return *(uint32_t*)&f32;
+}}
+
+static uint64_t pack_fp64(double f) {{
+    return *(uint64_t*)&f;
+}}
+
 static PyObject* launch(PyObject* self, PyObject* args) {{
   // ensure cuda context is valid before calling any CUDA APIs, e.g. before getPointer calls cuPointerGetAttributes
   ensureCudaContext();
@@ -462,14 +529,16 @@ static PyObject* launch(PyObject* self, PyObject* args) {{
   uint64_t _stream;
   uint64_t _function;
   int launch_cooperative_grid;
+  int launch_pdl;
   PyObject *launch_enter_hook = NULL;
   PyObject *launch_exit_hook = NULL;
   PyObject *kernel_metadata = NULL;
   PyObject *launch_metadata = NULL;
   PyObject *global_scratch_obj = NULL;
+  PyObject *profile_scratch_obj = NULL;
   {newline.join([f"{_extracted_type(ty)} _arg{i};" for i, ty in signature.items()])}
   if(!PyArg_ParseTuple(args, \"{format}\", &gridX, &gridY, &gridZ,
-                                           &_stream, &_function, &launch_cooperative_grid, &global_scratch_obj,
+                                           &_stream, &_function, &launch_cooperative_grid, &launch_pdl, &global_scratch_obj, &profile_scratch_obj,
                                            &kernel_metadata, &launch_metadata,
                                            &launch_enter_hook, &launch_exit_hook{args_list})) {{
     return NULL;
@@ -483,11 +552,10 @@ static PyObject* launch(PyObject* self, PyObject* args) {{
 
   // extract launch metadata
   if (launch_enter_hook != Py_None){{
-    PyObject* args = Py_BuildValue("(O)", launch_metadata);
-    PyObject* ret = PyObject_CallObject(launch_enter_hook, args);
-    Py_DECREF(args);
+    PyObject* ret = PyObject_CallOneArg(launch_enter_hook, launch_metadata);
     if (!ret)
       return NULL;
+    Py_DECREF(ret);
   }}
 
   CUdeviceptr global_scratch = 0;
@@ -499,23 +567,31 @@ static PyObject* launch(PyObject* self, PyObject* args) {{
     global_scratch = global_scratch_info.dev_ptr;
   }}
 
+  CUdeviceptr profile_scratch = 0;
+  if (profile_scratch_obj != Py_None) {{
+    DevicePtrInfo profile_scratch_info = getPointer(profile_scratch_obj, -1);
+    if (!profile_scratch_info.valid) {{
+      return NULL;
+    }}
+    profile_scratch = profile_scratch_info.dev_ptr;
+  }}
+
   // raise exception asap
   {newline.join(ptr_decls)}
   {newline.join(tma_decls)}
+  {newline.join(float_storage_decls)}
   Py_BEGIN_ALLOW_THREADS;
-  _launch(gridX, gridY, gridZ, num_warps, num_ctas, launch_cooperative_grid, clusterDimX, clusterDimY, clusterDimZ, shared_memory, (CUstream)_stream, (CUfunction)_function, global_scratch{', ' + ', '.join(internal_args_list) if len(internal_args_list) > 0 else ''});
+  _launch(gridX, gridY, gridZ, num_warps, num_ctas, launch_cooperative_grid, launch_pdl, clusterDimX, clusterDimY, clusterDimZ, shared_memory, (CUstream)_stream, (CUfunction)_function, global_scratch, profile_scratch{', ' + ', '.join(internal_args_list) if len(internal_args_list) > 0 else ''});
   Py_END_ALLOW_THREADS;
   if (PyErr_Occurred()) {{
     return NULL;
   }}
 
   if(launch_exit_hook != Py_None){{
-    PyObject* args = Py_BuildValue("(O)", launch_metadata);
-    PyObject* ret = PyObject_CallObject(launch_exit_hook, args);
-    Py_DECREF(args);
+    PyObject* ret = PyObject_CallOneArg(launch_exit_hook, launch_metadata);
     if (!ret)
       return NULL;
-
+    Py_DECREF(ret);
   }}
 
   Py_RETURN_NONE;
@@ -535,6 +611,19 @@ static struct PyModuleDef ModuleDef = {{
 }};
 
 PyMODINIT_FUNC PyInit___triton_launcher(void) {{
+  data_ptr_str = PyUnicode_InternFromString("data_ptr");
+  if(data_ptr_str == NULL) {{
+    return NULL;
+  }}
+  PyObject* driver_mod = PyImport_ImportModule("triton.backends.nvidia.driver");
+  if (driver_mod == NULL) {{
+    return NULL;
+  }}
+  py_tensor_map_type = PyObject_GetAttrString(driver_mod, "PyCUtensorMap");
+  if (py_tensor_map_type == NULL) {{
+    return NULL;
+  }}
+
   PyObject *m = PyModule_Create(&ModuleDef);
   if(m == NULL) {{
     return NULL;
@@ -546,6 +635,77 @@ PyMODINIT_FUNC PyInit___triton_launcher(void) {{
     return src
 
 
+# The TMA dtype enum values are slightly different on host vs device...
+TMA_DTYPE_DEVICE_TO_HOST = dict((i, i) for i in range(16))
+TMA_DTYPE_DEVICE_TO_HOST[8] = 10
+TMA_DTYPE_DEVICE_TO_HOST[9] = 8
+TMA_DTYPE_DEVICE_TO_HOST[10] = 9
+
+
+def make_tensordesc_arg(arg, metadata):
+    if metadata is None:
+        # Currently the host side tensor descriptors get decomposed in
+        # the frontend to tensor desc, shape, and strides. We have no
+        # way to use these shape and strides when processing tensor
+        # descriptors which is why we provide our own decomposition
+        # above. Sadly this means we have to pass the shape and strides
+        # twice.
+        return [arg.base, *arg.shape, *arg.strides, arg.padding == "nan", *arg.shape, *arg.strides]
+
+    swizzle = metadata["swizzle"]
+    elem_size = metadata["elem_size"]
+    elem_type = metadata["elem_type"]
+    block_size = metadata["block_size"]
+    fp4_padded = metadata["fp4_padded"]
+
+    shape = arg.shape
+    strides = arg.strides
+    assert strides[-1] == 1
+    padding = 1 if arg.padding == "nan" else 0
+
+    if fp4_padded:
+        shape = list(shape)
+        shape[-1] *= 2
+
+    cu_tensor_map = triton.runtime.driver.active.utils.fill_tma_descriptor(
+        arg.base.data_ptr(),
+        swizzle,
+        elem_size,
+        TMA_DTYPE_DEVICE_TO_HOST[elem_type],
+        block_size,
+        shape,
+        strides,
+        padding,
+    )
+
+    return [cu_tensor_map, *shape, *strides]
+
+
+def wrap_handle_tensordesc(launcher, signature, tensordesc_meta):
+    has_tensor_desc_arg = any(isinstance(sig, str) and sig.startswith("tensordesc") for sig in signature.values())
+    if not has_tensor_desc_arg:
+        return launcher
+
+    tensordesc_indices = set(
+        [i for i, sig in enumerate(signature.values()) if isinstance(sig, str) and sig.startswith("tensordesc")])
+    assert not tensordesc_meta or len(tensordesc_meta) == len(tensordesc_indices)
+    if not tensordesc_meta:
+        tensordesc_meta = [None] * len(tensordesc_indices)
+
+    def inner(*args):
+        final_args = list(args[:_BASE_ARGS_FORMAT_LEN])
+        tensordesc_idx = 0
+        for i, arg in enumerate(args[_BASE_ARGS_FORMAT_LEN:]):
+            if i in tensordesc_indices:
+                final_args.extend(make_tensordesc_arg(arg, tensordesc_meta[tensordesc_idx]))
+                tensordesc_idx += 1
+            else:
+                final_args.append(arg)
+        return launcher(*final_args)
+
+    return inner
+
+
 class CudaLauncher(object):
 
     def __init__(self, src, metadata):
@@ -553,21 +713,40 @@ class CudaLauncher(object):
         arg_idx = lambda x: (src.fn.arg_names.index(x), ) if isinstance(x, str) else x
         constants = {arg_idx(idx): value for idx, value in constants.items()}
         signature = {idx: value for idx, value in src.signature.items()}
-        src = make_launcher(constants, signature)
-        mod = compile_module_from_src(src, "__triton_launcher")
-        self.launch = mod.launch
+        tensordesc_meta = getattr(metadata, "tensordesc_meta", None)
+        src = make_launcher(constants, signature, tensordesc_meta)
+        mod = compile_module_from_src(
+            src=src,
+            name="__triton_launcher",
+            library_dirs=library_dirs(),
+            include_dirs=include_dirs,
+            libraries=libraries,
+        )
+
+        self.num_ctas = functools.reduce(operator.mul, metadata.cluster_dims, 1)
+        self.launch = wrap_handle_tensordesc(mod.launch, signature, tensordesc_meta)
         self.global_scratch_size = metadata.global_scratch_size
         self.global_scratch_align = metadata.global_scratch_align
+        self.profile_scratch_size = metadata.profile_scratch_size
+        self.profile_scratch_align = metadata.profile_scratch_align
         self.launch_cooperative_grid = metadata.launch_cooperative_grid
+        self.launch_pdl = metadata.launch_pdl
 
     def __call__(self, gridX, gridY, gridZ, stream, function, *args):
-        if self.global_scratch_size > 0:
-            grid_size = gridX * gridY * gridZ
-            alloc_size = grid_size * self.global_scratch_size
-            global_scratch = _allocation._allocator(alloc_size, self.global_scratch_align, stream)
-        else:
-            global_scratch = None
-        self.launch(gridX, gridY, gridZ, stream, function, self.launch_cooperative_grid, global_scratch, *args)
+
+        def allocate_scratch(size, align, allocator):
+            if size > 0:
+                grid_size = gridX * gridY * gridZ
+                alloc_size = grid_size * self.num_ctas * size
+                alloc_fn = allocator.get()
+                return alloc_fn(alloc_size, align, stream)
+            return None
+
+        global_scratch = allocate_scratch(self.global_scratch_size, self.global_scratch_align, _allocation._allocator)
+        profile_scratch = allocate_scratch(self.profile_scratch_size, self.profile_scratch_align,
+                                           _allocation._profile_allocator)
+        self.launch(gridX, gridY, gridZ, stream, function, self.launch_cooperative_grid, self.launch_pdl,
+                    global_scratch, profile_scratch, *args)
 
 
 class CudaDriver(GPUDriver):
@@ -600,6 +779,9 @@ class CudaDriver(GPUDriver):
         except ImportError:
             return False
 
+    def map_python_to_cpp_type(self, ty: str) -> str:
+        return ty_to_cpp(ty)
+
     def get_benchmarker(self):
         from triton.testing import do_bench
         return do_bench