Allow Kernels for Full FT and Non-Quantized PEFT (#79)

* add or logic for plugin registration Signed-off-by: Yu Chin Fabian Lim <[email protected]> Signed-off-by: 1000850000 user <[email protected]> * add fast kernels plugin Signed-off-by: Yu Chin Fabian Lim <[email protected]> Signed-off-by: 1000850000 user <[email protected]> * prepare full-foak benchmarks Signed-off-by: 1000850000 user <[email protected]> * update benchmark logic to have empty framework_config Signed-off-by: Yu Chin Fabian Lim <[email protected]> Signed-off-by: 1000850000 user <[email protected]> * minor fixes to foak full Signed-off-by: 1000850000 user <[email protected]> * addressed code review changes Signed-off-by: 1000850000 user <[email protected]> * additional fixes from code review Signed-off-by: 1000850000 user <[email protected]> * minor fixes to standard peft Signed-off-by: 1000850000 user <[email protected]> * Apply suggestions from code review Co-authored-by: Yu Chin Fabian Lim <[email protected]> Signed-off-by: 1000850000 user <[email protected]> * changes to filtering function and modifications to allow flexibilty of activating kernels Signed-off-by: 1000850000 user <[email protected]> * additional check in fastkernels and changes to FOAK README.md Signed-off-by: 1000850000 user <[email protected]> * fix syntax error Signed-off-by: 1000850000 user <[email protected]> * fix reloads on multiple patches Signed-off-by: 1000850000 user <[email protected]> * dtype changes to scenarios.yaml and README.md Signed-off-by: 1000850000 user <[email protected]> * changes to scenarios.yaml Signed-off-by: 1000850000 user <[email protected]> * additional comments Signed-off-by: 1000850000 user <[email protected]> * format and lint Signed-off-by: 1000850000 user <[email protected]> * fixes and updates to benchmark Signed-off-by: 1000850000 user <[email protected]> * fixes on reload Signed-off-by: Yu Chin Fabian Lim <[email protected]> --------- Signed-off-by: Yu Chin Fabian Lim <[email protected]> Signed-off-by: 1000850000 user <[email protected]> Co-authored-by: 1000850000 user <[email protected]> Co-authored-by: achew010 <[email protected]>
foundation-model-stack · Sep 16, 2024 · 4e81c64 · 4e81c64
1 parent a0ac97a
commit 4e81c64
Show file tree

Hide file tree

Showing 25 changed files with 890 additions and 179 deletions.
diff --git a/plugins/accelerated-peft/src/fms_acceleration_peft/framework_plugin_autogptq.py b/plugins/accelerated-peft/src/fms_acceleration_peft/framework_plugin_autogptq.py
@@ -231,7 +231,6 @@ def model_loader(self, model_name: str, **kwargs):
         #   and there is a section of code that will be skipped if not set.
         setattr(model, "is_loaded_in_4bit", True)
         setattr(model, "quantization_method", "gptq")
-
         return model
 
     @property
@@ -275,6 +274,8 @@ def augmentation(
 
         # some assertions
         assert peft_config is not None, "need peft_config to install PEFT adapters"
+        # running this plugin in float16 is the most performant
+        # https://github.com/foundation-model-stack/fms-acceleration/issues/84
         assert (
             model.dtype == torch.float16 or train_args.fp16
         ), "need to run in fp16 mixed precision or load model in fp16"
@@ -324,6 +325,13 @@ def augmentation(
             auto_find_all_linears=requires_installation_on_all_linears(peft_config),
             train_mode=True,  # install adapaters for training
         )
+
+        # We do not set `is_loaded_in_4bit`` at this point because otherwise
+        # `accelerate.prepare_model` will think the device placement is finalized
+        # for the quantized model, and will raise
+        # Reassign `quantization_method` after PEFT installation replaces the top-level class
+        setattr(model, "quantization_method", "gptq")
+
         modifiable_args = (None,)  # return a None for peft_config
 
         if self.use_external_lib:

diff --git a/...ention-and-distributed-packing/src/fms_acceleration_aadp/framework_plugin_padding_free.py b/...ention-and-distributed-packing/src/fms_acceleration_aadp/framework_plugin_padding_free.py
@@ -175,13 +175,20 @@ def _is_backbone(module: torch.nn.Module):
             # Local
             from .flash_attn import _flash_attention_forward_with_posids
 
+            # - we need to reload on the correct module
+            try:
+                # if it is peft
+                _module_path = model.get_base_model().__module__
+            except AttributeError:
+                _module_path = model.__module__
+
             ModelPatcher.register(
                 ModelPatcherRule(
                     rule_id="flash_attn_forward",
                     import_and_maybe_reload=(
                         "transformers.modeling_flash_attention_utils._flash_attention_forward",
                         partial(_flash_attention_forward_with_posids, id(model)),
-                        model.__module__,
+                        _module_path,
                     ),
                 ),
             )

diff --git a/plugins/framework/pyproject.toml b/plugins/framework/pyproject.toml
@@ -24,7 +24,6 @@ classifiers=[
 dependencies = [
   "numpy<2.0", # numpy needs to be bounded due to incompatiblity with current torch<2.3
   "torch>2.2",
-  "transformers",
   "peft",
   "accelerate",
   "pandas",

diff --git a/plugins/framework/src/fms_acceleration/framework_plugin.py b/plugins/framework/src/fms_acceleration/framework_plugin.py
@@ -29,7 +29,7 @@
 class PluginRegistration:
     plugin: "AccelerationPlugin"
     AND: List[str] = None
-    # OR: List[str] = None # not implemented yet
+    OR: List[str] = None
 
     # package metadata
     package_name: str = None
@@ -53,28 +53,61 @@ def _trace_key_path(configuration: Dict, key: str):
 def get_relevant_configuration_sections(configuration: Dict) -> Dict:
     results = []
 
+    # this function updates cfg with content
+    # - equivalent to taking a union
+    def _update_config_contents(_cfg: Dict, content: Dict, key: str):
+        path = key.split(".")
+        n = len(path)
+        _cfg = relevant_config
+        while n > 1:
+            p = path.pop(0)
+            if p not in _cfg:
+                _cfg[p] = {}
+            _cfg = _cfg[p]
+            n -= 1
+
+        _cfg[path[0]] = content
+
     # assume the registrations are all done with at least some default key
     for registration in PLUGIN_REGISTRATIONS:
         relevant_config = {}
-        # OR is not implemented yet
+
+        _and_keys = registration.AND
+        _or_keys = registration.OR
+        if _and_keys is None:
+            _and_keys = []
+        if _or_keys is None:
+            _or_keys = []
+
+        # go through AND paths then OR paths
+        # - if all AND paths are speciied, then return their union of all content
+        # - if any OR path is specified, then return the union of specified content
         reject = False
-        for key in registration.AND:
+        for key in _and_keys:
             content = _trace_key_path(configuration, key)
             if content is None:
+                # if AND key, then if at least one of them not
+                # specified, then reject and do not descend config tree
                 reject = True
                 break
 
-            path = key.split(".")
-            n = len(path)
-            _cfg = relevant_config
-            while n > 1:
-                p = path.pop(0)
-                if p not in _cfg:
-                    _cfg[p] = {}
-                _cfg = _cfg[p]
-                n -= 1
+            # update
+            _update_config_contents(relevant_config, content, key)
+
+        # if all the any keys were not satisfied, then reset the config
+        if reject:
+            relevant_config = {}
+
+        for key in _or_keys:
+            content = _trace_key_path(configuration, key)
+            if content is not None:
+                if reject:
+                    # it is an OR key, and if at least one of them specified
+                    # then do not reject
+                    reject = False
 
-            _cfg[path[0]] = content
+                # update all content that is not None
+                _update_config_contents(relevant_config, content, key)
 
         if reject:
             continue
@@ -91,7 +124,8 @@ class AccelerationPlugin:
     @staticmethod
     def register_plugin(
         plugin: "AccelerationPlugin",
-        configuration_and_paths: List[str],
+        configuration_and_paths: List[str] = None,
+        configuration_or_paths: List[str] = None,
         **kwargs,
     ):
 
@@ -101,6 +135,12 @@ def register_plugin(
         # is done (global-variable-not-assigned)
         # global PLUGIN_REGISTRATIONS
 
+        assert (
+            configuration_and_paths is not None and len(configuration_and_paths) > 0
+        ) or (
+            configuration_or_paths is not None and len(configuration_or_paths) > 0
+        ), "Specify at least one AND or OR path"
+
         # get the package metadata
         pkg_name = sys.modules[plugin.__module__].__package__
         try:
@@ -112,6 +152,7 @@ def register_plugin(
             PluginRegistration(
                 plugin=plugin,
                 AND=configuration_and_paths,
+                OR=configuration_or_paths,
                 package_name=pkg_name,
                 package_version=package_version,
             )

diff --git a/plugins/framework/src/fms_acceleration/model_patcher.py b/plugins/framework/src/fms_acceleration/model_patcher.py
@@ -348,14 +348,24 @@ def _import_and_reload(model: torch.nn.Module):
                 key=lambda _rule: len(_rule.import_and_maybe_reload[2]),
                 reverse=False,
             )
-            for rule_s in _with_reload:
-                for rule_l in _with_reload[1:]:
+
+            for i_s, rule_s in enumerate(_with_reload[:-1]):
+                for rule_l in _with_reload[i_s + 1 :]:
                     # if target paths in rule s is a prefix of rule l, raise an error
-                    _, _, _path_s = rule_s.import_and_maybe_reload
+                    _name_s, _obj_s, _path_s = rule_s.import_and_maybe_reload
                     _, _, _path_l = rule_l.import_and_maybe_reload
+
+                    if _path_s == _path_l:
+                        # - in the even the target is exactly the same, we will
+                        # only reload once
+                        rule_s.import_and_maybe_reload = (_name_s, _obj_s, None)
+                        continue
+
+                    # - otherwise, we do not consider the cases where the target
+                    # is a subpath since this results in unpredictablity.
                     assert not _path_l.startswith(
                         _path_s
-                    ), f"Attempting to reload same path `{_path_s}` multiple times in \
+                    ), f"Attempting to reload a subpath`{_path_s}` multiple times in \
                             {rule_s.rule_id} and {rule_l.rule_id}"
 
         # handle those with reload first

diff --git a/plugins/framework/src/fms_acceleration/utils/test_utils.py b/plugins/framework/src/fms_acceleration/utils/test_utils.py
@@ -18,7 +18,7 @@
 # Standard
 from contextlib import contextmanager
 from tempfile import NamedTemporaryFile
-from typing import Any, Callable, Dict, List, Set, Tuple, Type
+from typing import Any, Callable, Dict, List, Set, Tuple, Type, Union
 
 # Third Party
 import torch
@@ -67,7 +67,14 @@ def configure_framework_from_json(
 @contextmanager
 def build_framework_and_maybe_instantiate(
     plugins_to_be_registered: List[
-        Tuple[List[str], Type[AccelerationPlugin]]  # and_paths, plugin_class
+        Union[
+            Tuple[List[str], Type[AccelerationPlugin]],  # and_paths, plugin_class
+            Tuple[
+                List[str],
+                List[str],  # and_or_paths
+                Type[AccelerationPlugin],  # plugin_class
+            ],
+        ]
     ],
     configuration_contents: Dict = None,
     instantiate: bool = True,
@@ -89,10 +96,17 @@ def build_framework_and_maybe_instantiate(
     AccelerationFramework.active_plugins = []
     AccelerationFramework.plugins_require_custom_loading = []
 
-    for path, plugin in plugins_to_be_registered:
+    for paths_and_plugins in plugins_to_be_registered:
+        try:
+            and_paths, plugin = paths_and_plugins
+            or_paths = None
+        except ValueError:
+            and_paths, or_paths, plugin = paths_and_plugins
+
         AccelerationPlugin.register_plugin(
             plugin,
-            configuration_and_paths=path,
+            configuration_and_paths=and_paths,
+            configuration_or_paths=or_paths,
         )
 
     if instantiate:

diff --git a/plugins/framework/tests/test_framework.py b/plugins/framework/tests/test_framework.py
@@ -313,3 +313,76 @@ def _hook(
         framework.augmentation(model, None, None)
         for c, (n, _) in zip(plugin_activation_order, plugins_to_be_installed):
             assert n in c
+
+
+def test_plugin_registration_combination_logic():
+
+    plugin = create_plugin_cls(
+        restricted_models={"CausalLM"},
+        requires_agumentation=True,
+        agumentation=dummy_augmentation,
+    )
+
+    configuration_contents = {"existing1": {"key1": 1}, "existing2": {"key1": 1}}
+
+    # empty conditions
+    with pytest.raises(AssertionError, match="Specify at least one AND or OR path"):
+        with build_framework_and_instantiate(
+            plugins_to_be_registered=[
+                ([], [], plugin),
+            ],
+            configuration_contents=configuration_contents,
+        ) as framework:
+            pass
+
+    # AND logic - happy
+    with build_framework_and_instantiate(
+        plugins_to_be_registered=[
+            (["existing1", "existing2"], plugin),
+        ],
+        configuration_contents=configuration_contents,
+    ) as framework:
+        # check 1.
+        assert len(PLUGIN_REGISTRATIONS) == 1
+
+        # check 2.
+        assert len(framework.active_plugins) == 1
+
+    # AND  - sad path
+    with pytest.raises(
+        ValueError,
+        match="No plugins could be configured. Please check the acceleration",
+    ):
+        with build_framework_and_instantiate(
+            plugins_to_be_registered=[
+                (["existing1", "non-existant"], plugin),
+            ],
+            configuration_contents=configuration_contents,
+        ) as framework:
+            pass
+
+    # OR logic
+    with build_framework_and_instantiate(
+        plugins_to_be_registered=[
+            ([], ["existing1", "non-existant"], plugin),
+        ],
+        configuration_contents=configuration_contents,
+    ) as framework:
+        # check 1.
+        assert len(PLUGIN_REGISTRATIONS) == 1
+
+        # check 2.
+        assert len(framework.active_plugins) == 1
+
+    # OR - sad path
+    with pytest.raises(
+        ValueError,
+        match="No plugins could be configured. Please check the acceleration",
+    ):
+        with build_framework_and_instantiate(
+            plugins_to_be_registered=[
+                (["non-existant", "non-existant2"], plugin),
+            ],
+            configuration_contents=configuration_contents,
+        ) as framework:
+            pass