Pipeline extraction

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>

Author: kylesayrs

src/llmcompressor/args/dataset_arguments.py

@@ -3,6 +3,8 @@
 
 from transformers import DefaultDataCollator
 
+from llmcompressor.pipelines.registry import PIPELINES
+
 
 @dataclass
 class DVCDatasetArguments:
@@ -171,3 +173,10 @@ class DatasetArguments(CustomDatasetArguments):
             "will execute code present on the Hub on your local machine."
         },
     )
+    pipeline: Optional[str] = field(
+        default="independent",
+        metadata={
+            "help": "Calibration pipeline used to calibrate model. "
+            f"Options: {PIPELINES.keys()}"
+        },
+    )

src/llmcompressor/core/events/event.py

@@ -44,6 +44,8 @@ class EventType(Enum):
     BATCH_START = "batch_start"
     LOSS_CALCULATED = "loss_calculated"
     BATCH_END = "batch_end"
+    SEQUENTIAL_EPOCH_END = "sequential_epoch_end"
+    CALIBRATION_EPOCH_END = "calibration_epoch_end"
 
     # step lifecycle
     OPTIM_PRE_STEP = "optim_pre_step"

src/llmcompressor/core/lifecycle.py

@@ -77,6 +77,16 @@ def reset(self):
         self.__init__()
         logger.info("Compression lifecycle reset")
 
+    def initialize_recipe(
+        self,
+        recipe: Optional[RecipeInput] = None,
+        recipe_stage: Optional[RecipeStageInput] = None,
+        recipe_args: Optional[RecipeArgsInput] = None,
+    ):
+        if len(self.modifiers) <= 0:
+            self.recipe_container.append(recipe, recipe_stage, recipe_args)
+            self.modifiers = self.recipe_container.get_modifiers()
+
     def initialize(
         self,
         recipe: Optional[RecipeInput] = None,
@@ -92,12 +102,9 @@ def initialize(
         :rtype: List[Any]
         """
         self.state.update(**kwargs)
-        if self.initialized_:  # TODO: do not initialize twice
-            return
 
         logger.debug("Initializing compression lifecycle")
-        self.recipe_container.append(recipe, recipe_stage, recipe_args)
-        self.modifiers = self.recipe_container.get_modifiers()
+        self.initialize_recipe(recipe, recipe_stage, recipe_args)
         self._set_model_layer_prefix()
 
         mod_data = []

src/llmcompressor/core/session.py

@@ -1,4 +1,5 @@
 from dataclasses import dataclass
+from functools import reduce
 from typing import Any, Callable, Dict, List, Optional, Union
 
 from loguru import logger
@@ -220,6 +221,10 @@ def get_serialized_recipe(self) -> Optional[str]:
 
         logger.warning("Recipe not found in session - it may have been reset")
 
+    def get_modifiers(self):
+        stage_modifiers = self.lifecycle.modifiers
+        return list(reduce(sum, (mod.modifiers for mod in stage_modifiers)))
+
     def _log_model_info(self):
         # Log model level logs if cadence reached
         current_index = self._lifecycle.global_step

src/llmcompressor/core/session_functions.py

@@ -1,6 +1,6 @@
 import threading
 from contextlib import contextmanager
-from typing import Any, Optional
+from typing import Any, Generator, Optional
 
 from llmcompressor.core.events import EventType
 from llmcompressor.core.session import CompressionSession
@@ -21,7 +21,7 @@
 
 
 @contextmanager
-def create_session() -> CompressionSession:
+def create_session() -> Generator[CompressionSession, None, None]:
     """
     Context manager to create and yield a new session for sparsification.
     This will set the active session to the new session for the duration
@@ -136,5 +136,26 @@ def batch_end(cls, **kwargs) -> ModifiedState:
         active_session()._log_model_info()
         return cls.event(EventType.BATCH_END, **kwargs)
 
+    @classmethod
+    def sequential_epoch_end(cls, **kwargs) -> ModifiedState:
+        """
+        Invoke a sequential epoch end event for the active session. This event should be
+        called after one sequential layer has been calibrated/trained for one epoch
+
+        This is called after a sequential layer has been calibrated with one batch, see
+        `src/llmcompressor/pipelines/sequential/pipeline.py` for usage example
+        """
+        return cls.event(EventType.SEQUENTIAL_EPOCH_END, **kwargs)
+
+    @classmethod
+    def calibration_epoch_end(cls, **kwargs) -> ModifiedState:
+        """
+        Invoke a epoch end event for the active session during calibration. This event
+        should be called after the model has been calibrated for one epoch
+
+        see `src/llmcompressor/pipelines/basic/pipeline.py` for usage example
+        """
+        return cls.event(EventType.CALIBRATION_EPOCH_END, **kwargs)
+
 
 callbacks = LifecycleCallbacks

src/llmcompressor/entrypoints/oneshot.py

@@ -7,6 +7,7 @@
 from llmcompressor.core.session_functions import active_session
 from llmcompressor.datasets import get_calibration_dataloader
 from llmcompressor.entrypoints.utils import post_process, pre_process
+from llmcompressor.pipelines.registry import get_pipeline_fn
 
 __all__ = ["Oneshot", "oneshot"]
 
@@ -157,21 +158,20 @@ def apply_recipe_modifiers(
         """
 
         session = active_session()
+        session.reset()
 
-        session_kwargs = dict(
-            model=self.model,
+        session.lifecycle.state.update(model=self.model, start=-1)
+        session.lifecycle.initialize_recipe(
             recipe=self.recipe,
-            recipe_args=self.recipe_args.recipe_args,
-            calib_data=calibration_dataloader,
-            start=-1,  # oneshot-specific argument
-            copy_data=False,
-            min_tokens_per_module=getattr(self, "min_tokens_per_module", None),
             recipe_stage=recipe_stage,
+            recipe_args=self.recipe_args.recipe_args,
         )
 
-        session.reset()
-        session.initialize(**session_kwargs)
-        session.finalize(**session_kwargs)
+        modifiers = session.get_modifiers()
+        _, pipeline_fn = get_pipeline_fn(self.dataset_args.pipeline, modifiers)
+        pipeline_fn(self.model, calibration_dataloader)
+
+        session.finalize()
 
 
 def oneshot(**kwargs) -> PreTrainedModel:

src/llmcompressor/modifiers/obcq/base.py

@@ -10,7 +10,7 @@
 from loguru import logger
 from pydantic import PrivateAttr
 
-from llmcompressor.core import State
+from llmcompressor.core import Event, EventType, State
 from llmcompressor.modifiers import Modifier
 from llmcompressor.modifiers.obcq.sgpt_mixin import SparsityModifierMixin
 from llmcompressor.modifiers.obcq.sgpt_sparsify import (
@@ -90,6 +90,14 @@ def calibrate_module(
         args: Tuple[torch.Tensor, ...],
         _output: torch.Tensor,
     ):
+        """
+        Calibration hook used to accumulate the hessian of the input to the module
+
+        :param module: module being calibrated
+        :param args: inputs to the module, the first element of which is the
+            cannonical input
+        :param _output: uncompressed module output, unused
+        """
         # Assume that the first argument is the input
         inp = args[0]
 
@@ -108,10 +116,16 @@ def calibrate_module(
                 self._num_samples[module],
             )
 
-    def on_sequential_batch_end(self):
+    def on_event(self, state: State, event: Event, **kwargs):
+        if event.type_ in (
+            EventType.SEQUENTIAL_EPOCH_END,
+            EventType.CALIBRATION_EPOCH_END,
+        ):
+            self.compress_modules()
+
+    def compress_modules(self):
         """
-        Sparsify modules
-        TODO: implement with event callback
+        Sparsify modules which have been calibrated
         """
         for module in list(self._num_samples.keys()):
             name = self._module_names[module]
@@ -152,6 +166,9 @@ def _maybe_onload_hessian(self, module: torch.nn.Module):
                 self._hessians[module] = self._hessians[module].to(device="cpu")
 
     def on_finalize(self, state: State, **kwargs) -> bool:
+        if len(self._num_samples) > 0:
+            raise ValueError(f"Failed to compress {len(self._num_samples)} modules")
+
         self.remove_hooks()
         self._hessians = dict()
         self._num_samples = dict()

src/llmcompressor/modifiers/obcq/sgpt_mixin.py

@@ -12,10 +12,6 @@
 from llmcompressor.core import State
 from llmcompressor.modifiers.utils.hooks import HooksMixin
 from llmcompressor.pipelines.basic import run_pipeline as run_basic
-from llmcompressor.pipelines.layer_sequential import (
-    run_pipeline as run_layer_sequential,
-)
-from llmcompressor.pipelines.sequential import run_pipeline as run_sequential
 from llmcompressor.utils.pytorch.module import (
     get_layers,
     get_no_split_params,
@@ -160,49 +156,6 @@ def on_initialize(self, state: "State", **kwargs) -> bool:
                 self._module_sparsities[module] = layer_sparsity
                 self.register_hook(module, self.calibrate_module, "forward")
 
-        # infer and run pipeline
-        model_name = state.model.__class__.__name__
-        input_names = dataloader.dataset.column_names
-        unfixable_errors = (torch.OutOfMemoryError, torch._C._LinAlgError)
-        try:
-            run_sequential(
-                state.model,
-                state.data.calib,
-                self.sequential_targets,
-                self.ignore,
-                self,
-            )
-            return True
-
-        except Exception as exception:
-            if isinstance(exception, torch.fx.proxy.TraceError):
-                warnings.warn(f"Failed to trace {model_name} with inputs {input_names}")
-            if isinstance(exception, unfixable_errors):
-                raise exception
-
-            warnings.warn("Falling back to layer_sequential pipeline")
-            try:
-                run_layer_sequential(
-                    state.model,
-                    state.data.calib,
-                    self.sequential_targets,
-                    self,
-                )
-                return True
-
-            except Exception as exception:
-                if isinstance(exception, TypeError):
-                    warnings.warn(f"{model_name} fails layer-wise assumptions")
-                if isinstance(exception, unfixable_errors):
-                    raise exception
-
-                warnings.warn(
-                    "Falling back to basic pipeline, which requires extra memory and "
-                    "may result in decreased accuracy"
-                )
-                run_basic(state.model, state.data.calib, self)
-                return True
-
     def _infer_sequential_targets(
         self, model: torch.nn.Module
     ) -> Union[str, List[str]]:

src/llmcompressor/modifiers/pruning/wanda/base.py

@@ -9,7 +9,7 @@
 from loguru import logger
 from pydantic import PrivateAttr
 
-from llmcompressor.core import State
+from llmcompressor.core import Event, EventType, State
 from llmcompressor.modifiers import Modifier
 from llmcompressor.modifiers.obcq.sgpt_mixin import SparsityModifierMixin
 from llmcompressor.modifiers.pruning.wanda.wanda_sparsify import (
@@ -74,6 +74,14 @@ def calibrate_module(
         args: Tuple[torch.Tensor, ...],
         _output: torch.Tensor,
     ):
+        """
+        Calibration hook used to accumulate the row scalars of the input to the module
+
+        :param module: module being calibrated
+        :param args: inputs to the module, the first element of which is the
+            cannonical input
+        :param _output: uncompressed module output, unused
+        """
         # Assume that the first argument is the input
         inp = args[0]
 
@@ -91,12 +99,17 @@ def calibrate_module(
             self._num_samples[module],
         )
 
-    def on_sequential_batch_end(self):
+    def on_event(self, state: State, event: Event, **kwargs):
+        if event.type_ in (
+            EventType.SEQUENTIAL_EPOCH_END,
+            EventType.CALIBRATION_EPOCH_END,
+        ):
+            self.compress_modules()
+
+    def compress_modules(self):
         """
-        Sparsify modules
-        TODO: implement with event callback
+        Sparsify modules which have been calibrated
         """
-
         for module in list(self._num_samples.keys()):
             name = self._module_names[module]
             sparsity = self._module_sparsities[module]
@@ -120,6 +133,9 @@ def on_sequential_batch_end(self):
             del self._num_samples[module]
 
     def on_finalize(self, state: State, **kwargs) -> bool:
+        if len(self._num_samples) > 0:
+            raise ValueError(f"Failed to compress {len(self._num_samples)} modules")
+
         self.remove_hooks()
         self._row_scalars = dict()
         self._num_samples = dict()