[Contrib][Openfold Triton] Use json instead of pickle (#1900)

Aidyn-A · crcrpar · web-flow · commit e74a67bba3ee · 2025-05-16T01:17:16.000+09:00
* use json instead of pickle

* Remove space

* Add backward call

* Fix function arguments

* Add run_tests()

* Reorder imports

* Apply nit-picks

Co-Authored-By: Masaki Kozuki &lt;16191443+crcrpar@users.noreply.github.com&gt;

---------

Co-authored-by: Masaki Kozuki &lt;16191443+crcrpar@users.noreply.github.com&gt;
diff --git a/apex/contrib/openfold_triton/__init__.py b/apex/contrib/openfold_triton/__init__.py
@@ -1,13 +1,14 @@
 # © 2023 NVIDIA CORPORATION & AFFILIATES
 
-import pickle
+import json
+import warnings
 from collections import OrderedDict
 from copy import deepcopy
 from io import BytesIO
 from typing import BinaryIO, Union
 
 import torch
-from triton.runtime.autotuner import Autotuner, Heuristics
+from triton.runtime.autotuner import Autotuner, Config, Heuristics
 from triton.runtime.jit import JITFunction
 
 from apex.contrib.openfold_triton._layer_norm_backward_kernels import (
@@ -58,23 +59,27 @@ def _get_tuneable_triton_func_name(f: Union[Autotuner, Heuristics, JITFunction])
 )
 
 
-def _save_triton_auto_tune_cache(f: BinaryIO, verbose: bool = False) -> None:
+def _save_triton_auto_tune_cache(strict: bool = True, verbose: bool = False) -> BytesIO:
     caches = OrderedDict()
     for func_name, func in _tuneable_triton_kernels.items():
         if len(func.cache) < 1:
-            raise ValueError(
-                f"Triton JIT kernel {func.__name__} didn't have tuning cache"
-            )
-        caches[func_name] = deepcopy(func.cache)
-    pickle.dump(caches, f)
+            msg = f"Triton JIT kernel {func_name} didn't have tuning cache"
+            if strict:
+                raise ValueError(msg)
+            else:
+                warnings.warn(msg)
+        else:
+            caches[func_name] = [(keys, vals.all_kwargs()) for keys, vals in zip(func.cache.keys(), func.cache.values())]
+    f = BytesIO(json.dumps(caches).encode('utf-8'))
     if verbose:
         print(f"Triton kernel auto-tuning caches written to {f}")
+    return f
 
 
 def _load_triton_auto_tune_cache(
     f: BinaryIO, strict: bool = True, verbose: bool = False
 ) -> None:
-    caches = pickle.load(f)
+    caches = json.load(f)
     if strict:
         loaded_func_name = set(caches.keys())
         tuneable_func_name = set(_tuneable_triton_kernels.keys())
@@ -84,23 +89,23 @@ def _load_triton_auto_tune_cache(
                 f"Missing kernel caches: {tuneable_func_name - loaded_func_name}\n"
                 f"Unexpected kernel caches: {loaded_func_name - tuneable_func_name}"
             )
-    for func_name, cache in caches.items():
+    for func_name, func_cache in caches.items():
         if func_name not in _tuneable_triton_kernels:
             raise ValueError(
                 f"{func_name} from {f} doesn't match any tuneable Triton kernels"
             )
-        _tuneable_triton_kernels[func_name].cache = cache
+        for key, val in func_cache:
+            _tuneable_triton_kernels[func_name].cache[tuple(key)] = Config(val)
     if verbose:
         print(f"Triton kernel auto-tuning caches loaded from {f}")
 
 
-def sync_triton_auto_tune_cache_across_gpus() -> None:
+def sync_triton_auto_tune_cache_across_gpus(strict: bool = True, verbose: bool = False) -> None:
     if not torch.distributed.is_initialized():
         return
     if torch.distributed.get_rank() == 0:
         print("Broadcasting Triton auto-tuning cache from rank 0 to other ranks...")
-        cache = BytesIO()
-        _save_triton_auto_tune_cache(cache)
+        cache = _save_triton_auto_tune_cache(strict=strict, verbose=verbose)
         cache.seek(0)
         cache_list = [
             cache,
@@ -113,6 +118,5 @@ def sync_triton_auto_tune_cache_across_gpus() -> None:
             None,
         ]
     torch.distributed.broadcast_object_list(cache_list)
-    cache = cache_list[0]
-    _load_triton_auto_tune_cache(cache)
+    _load_triton_auto_tune_cache(cache_list[0], strict=strict, verbose=verbose)
     print("Succeed!")
diff --git a/apex/contrib/test/openfold_triton/test_sync_triton_auto_tune_cache_across_gpus.py b/apex/contrib/test/openfold_triton/test_sync_triton_auto_tune_cache_across_gpus.py
@@ -0,0 +1,97 @@
+import os
+import torch
+import torch.distributed as dist
+from torch.testing._internal.common_distributed import (
+    MultiProcessTestCase,
+    requires_nccl,
+    skip_if_lt_x_gpu,
+    run_tests,
+)
+from apex.contrib.openfold_triton import (
+    LayerNormSmallShapeOptImpl,
+    sync_triton_auto_tune_cache_across_gpus,
+    _tuneable_triton_kernels,
+)
+
+class SyncTritonAutoTuneCacheTest(MultiProcessTestCase):
+    device_type = "cuda"
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+
+    def setUp(self) -> None:
+        super().setUp()
+        self._spawn_processes()
+
+    def tearDown(self) -> None:
+        torch.cuda.synchronize()
+        torch.cuda.empty_cache()
+        super().tearDown()
+
+    @property
+    def world_size(self) -> int:
+        return min(torch.cuda.device_count(), 2)
+
+    @property
+    def init_method(self):
+        return f"{common_utils.FILE_SCHEMA}{self.file_name}"
+
+    @property
+    def destroy_pg_upon_exit(self) -> bool:
+        return True
+
+    def _create_process_group_nccl(self):
+        def maybe_export(env, val):
+            if not type(env) == str:
+                raise ValueError(f"Type of type of env is expected to be str, but got {type(env)}")
+            if not type(val) == str:
+                raise ValueError(f"Type of type of val is expected to be str, but got {type(val)}")
+            if os.getenv(env) is None:
+                os.environ[env] = val
+
+        maybe_export("MASTER_PORT", "29500")
+        maybe_export("MASTER_ADDR", "localhost")
+
+        # create nccl processgroup for two ranks
+        dist.init_process_group(
+            "nccl",
+            world_size=self.world_size,
+            rank=self.rank,
+        )
+        pg = dist.distributed_c10d._get_default_group()
+        return pg
+
+
+    @requires_nccl()
+    @skip_if_lt_x_gpu(1)
+    def test_sync_triton_auto_tune_cache_across_gpus(self):
+        pg = self._create_process_group_nccl()
+        device = torch.device(f"cuda:{self.rank % torch.cuda.device_count()}")
+        torch.cuda.set_device(device)
+
+        if self.rank == 0:
+            eps = 1e-5
+            normalized_shape = (128, 64,)
+
+            weight = torch.ones(normalized_shape, device=device, requires_grad=True)
+            bias= torch.zeros(normalized_shape, device=device, requires_grad=True)
+
+            x = torch.randn((2, 2,) + normalized_shape, device=device)
+            y = LayerNormSmallShapeOptImpl.apply(
+                x, normalized_shape, weight, bias, eps
+            )
+            l = torch.sum(y)
+            l.backward()
+
+        sync_triton_auto_tune_cache_across_gpus(strict=False, verbose=True)
+
+        caches_synced = 0
+        for func_name, func in _tuneable_triton_kernels.items():
+            if len(func.cache) > 0:
+                caches_synced = caches_synced + 1
+                print(f"caches were synchronized for {func_name} at rank = {self.rank}:", func.cache)
+
+        self.assertTrue(caches_synced > 0)
+
+
+if __name__ == '__main__':
+    run_tests()
