fusion-bench 0.2.17__py3-none-any.whl → 0.2.18__py3-none-any.whl

fusion_bench/models/expert_sparsity/mixtral/wrapper.py

@@ -0,0 +1,268 @@
+import itertools as I
+import logging
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers.models.mixtral.modeling_mixtral import (
+    MixtralBlockSparseTop2MLP,
+    MixtralForCausalLM,
+    MixtralSparseMoeBlock,
+)
+
+from .dataset import CacheDataset
+
+logger = logging.getLogger(__name__)
+
+
+class PrunableMixtralSparseMoeBlockWrapper(torch.nn.Module):
+    """
+    Wrapper of `MixtralSparseMoeBlock` that supports expert pruning.
+    """
+
+    def __init__(
+        self,
+        model: MixtralSparseMoeBlock,
+        r: Optional[int] = None,
+    ):
+        """
+        Args:
+            model: The model to be wrapped.
+            r: The number of experts to keep.
+        """
+        super().__init__()
+        if isinstance(model, MixtralSparseMoeBlock):
+            self.model = model
+        else:
+            self.model = model.model
+        self.r = r
+
+        self.experts_to_drop = None
+        self.cache_space = CacheDataset()
+        self.cache_logits = False
+        self.cache_X = False
+        self.cache_Z = False
+
+    # Forward uses topk
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """ """
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        # router_logits: (batch * sequence_length, n_experts)
+        router_logits = self.model.gate(hidden_states)
+
+        if self.experts_to_drop is not None:
+            for e in self.experts_to_drop:
+                router_logits[:, e] = -float("inf")
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(
+            routing_weights, self.model.top_k, dim=-1
+        )
+        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        # we cast back to the input dtype
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim),
+            dtype=hidden_states.dtype,
+            device=hidden_states.device,
+        )
+
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = torch.nn.functional.one_hot(
+            selected_experts, num_classes=self.model.num_experts
+        ).permute(2, 1, 0)
+
+        # Loop over all available experts in the model and perform the computation on each expert
+        for expert_idx in range(self.model.num_experts):
+            expert_layer = self.model.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx])
+
+            if top_x.shape[0] == 0:
+                continue
+
+            # in torch it is faster to index using lists than torch tensors
+            top_x_list = top_x.tolist()
+            idx_list = idx.tolist()
+
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim)
+            current_hidden_states = (
+                expert_layer(current_state)
+                * routing_weights[top_x_list, idx_list, None]
+            )
+
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(
+                0, top_x, current_hidden_states.to(hidden_states.dtype)
+            )
+
+        if self.experts_to_drop is not None and (
+            self.cache_logits or self.cache_X or self.cache_Z
+        ):
+            logger.warn(
+                f"Already dropped {self.experts_to_drop} but still storing activations."
+            )
+        self.cache_space.append(
+            alpha=(router_logits if self.cache_logits else None),
+            X=(hidden_states if self.cache_X else None),
+            Z=(final_hidden_states if self.cache_Z else None),
+        )
+
+        final_hidden_states = final_hidden_states.reshape(
+            batch_size, sequence_length, hidden_dim
+        )
+
+        return final_hidden_states, router_logits
+
+    @torch.no_grad()
+    def enumerate(self):
+        # disable caching
+        self.cache_logits = False
+        self.cache_X = False
+        self.cache_Z = False
+        loss_history = dict()
+
+        with torch.inference_mode():
+            for dropped in I.combinations(
+                range(self.model.num_experts), self.model.num_experts - self.r
+            ):
+                self.experts_to_drop = dropped
+                loss = 0
+
+                for hidden_states, final_hidden_states in zip(
+                    self.cache_space.Xs, self.cache_space.Zs
+                ):
+                    hidden_states = hidden_states.to(
+                        device=self.model.gate.weight.data.device, non_blocking=True
+                    )
+                    final_hidden_states = final_hidden_states.to(
+                        dtype=torch.float64,
+                        device=self.model.gate.weight.data.device,
+                        non_blocking=True,
+                    )
+                    final_hidden_states_e, _ = self.forward(hidden_states.unsqueeze(0))
+                    # compute the |Z - Z_e|_2 L2 loss
+                    loss += torch.norm(
+                        final_hidden_states
+                        - final_hidden_states_e.squeeze(0).to(torch.float64)
+                    ).item()
+                loss_history[dropped] = loss
+
+        self.experts_to_drop = min(loss_history, key=loss_history.get)
+        return loss_history
+
+    @torch.no_grad()
+    def prune(self):
+        assert self.experts_to_drop is not None
+        assert len(self.experts_to_drop) == self.model.num_experts - self.r
+        del self.cache_space
+        self.cache_X = False
+        self.cache_Z = False
+
+        experts_to_reserve = sorted(
+            set(range(self.model.num_experts)) - set(self.experts_to_drop)
+        )
+
+        # create a new gate with the experts to reserve
+        gate_new = torch.nn.Linear(
+            in_features=self.model.gate.in_features,
+            out_features=self.r,
+            bias=False,
+            device=self.model.gate.weight.data.device,
+            dtype=torch.bfloat16,
+        )
+        gate_new.weight.data = self.model.gate.weight.data[list(experts_to_reserve)]
+        self.model.gate = gate_new
+
+        self.model.experts = torch.nn.ModuleList(
+            [self.model.experts[i] for i in experts_to_reserve]
+        )
+        self.model.num_experts = self.r
+
+
+class DynamicSkippingMixtralSparseMoeBlockWrapper(nn.Module):
+    def __init__(self, model: MixtralSparseMoeBlock, beta: float):
+        super().__init__()
+        assert isinstance(model, MixtralSparseMoeBlock)
+        assert model.top_k == 2
+        self.hidden_dim = model.hidden_dim
+        self.ffn_dim = model.ffn_dim
+        self.num_experts = model.num_experts
+        self.top_k = model.top_k
+        self.gate = model.gate
+        self.experts = model.experts
+
+        self.beta = beta
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """ """
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        # router_logits: (batch * sequence_length, n_experts)
+        router_logits = self.gate(hidden_states)
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(
+            routing_weights, self.top_k, dim=-1
+        )
+
+        # (batch * sequence_length)
+        mask_top1 = routing_weights[:, 1] < self.beta * routing_weights[:, 0]
+        routing_weights[mask_top1, 1] = 0
+
+        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        # we cast back to the input dtype
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim),
+            dtype=hidden_states.dtype,
+            device=hidden_states.device,
+        )
+
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        # (batch * sequence_length, self.top_k, n_experts)
+        expert_mask = torch.nn.functional.one_hot(
+            selected_experts, num_classes=self.num_experts
+        )
+
+        expert_mask[mask_top1, 1, :] = 0
+        expert_mask = expert_mask.permute(2, 1, 0)
+
+        # Loop over all available experts in the model and perform the computation on each expert
+        for expert_idx in range(self.num_experts):
+            expert_layer = self.experts[expert_idx]
+            top_x, indices = torch.where(expert_mask[expert_idx])
+
+            if indices.shape[0] == 0:
+                continue
+
+            # in torch it is faster to index using lists than torch tensors
+            indices_list = indices.tolist()
+            top_x_list = top_x.tolist()
+
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states[None, indices_list].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer(
+                current_state, routing_weights[indices_list, top_x_list, None]
+            )
+
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(
+                0, indices, current_hidden_states.to(hidden_states.dtype)
+            )
+        final_hidden_states = final_hidden_states.reshape(
+            batch_size, sequence_length, hidden_dim
+        )
+        return final_hidden_states, router_logits

fusion_bench/programs/fabric_fusion_program.py

@@ -296,13 +296,17 @@ class FabricModelFusionProgram(
             if hydra_output_dir is not None:
                 os.makedirs(self.log_dir, exist_ok=True)
                 try:
-                    os.symlink(
-                        hydra_output_dir,
-                        os.path.join(
-                            self.log_dir,
-                            "hydra_output_" + os.path.basename(hydra_output_dir),
-                        ),
-                        target_is_directory=True,
-                    )
+                    # if the system is windows, use the `mklink` command in "CMD" to create the symlink
+                    if os.name == "nt":
+                        os.system(f"mklink /J {os.path.abspath(os.path.join(self.log_dir, 'hydra_output_' + os.path.basename(hydra_output_dir)))} {os.path.abspath(hydra_output_dir)}")
+                    else:
+                        os.symlink(
+                            hydra_output_dir,
+                            os.path.join(
+                                self.log_dir,
+                                "hydra_output_" + os.path.basename(hydra_output_dir),
+                            ),
+                            target_is_directory=True,
+                        )
                 except OSError as e:
                     log.warning(f"Failed to create symbolic link: {e}")

fusion_bench/utils/__init__.py

@@ -2,14 +2,15 @@
 import importlib
 from typing import Iterable
 
-from . import data, functools, path
+from . import data, functools, path, pylogger
 from .cache_utils import *
 from .devices import *
 from .dtype import parse_dtype
 from .fabric import seed_everything_by_time
 from .instantiate_utils import instantiate, is_instantiable
+from .json import load_from_json, save_to_json
+from .lazy_state_dict import LazyStateDict
 from .misc import *
 from .packages import import_object
 from .parameters import *
 from .timer import timeit_context
-from .lazy_state_dict import LazyStateDict

fusion_bench/utils/dtype.py

@@ -13,6 +13,7 @@ from transformers.utils import (
 PRECISION_STR_TO_DTYPE: Dict[str, torch.dtype] = {
     "fp16": torch.float16,
     "float16": torch.float16,
+    "half": torch.float16,
     "bf16": torch.bfloat16,
     "bfloat16": torch.bfloat16,
     "float": torch.float32,
@@ -50,7 +51,7 @@ def parse_dtype(dtype: Optional[str]):
 
     dtype = dtype.strip('"')
     if dtype not in PRECISION_STR_TO_DTYPE:
-        raise ValueError(f"Unsupported dtype: {type(dtype)}")
+        raise ValueError(f"Unsupported dtype string: {dtype}")
 
     dtype = PRECISION_STR_TO_DTYPE[dtype]
     return dtype

fusion_bench/utils/fabric.py

@@ -1,17 +1,24 @@
 import time
+from typing import Optional
 
 import lightning as L
 
+from fusion_bench.utils.pylogger import getRankZeroLogger
 
-def seed_everything_by_time(fabric: L.Fabric):
+log = getRankZeroLogger(__name__)
+
+
+def seed_everything_by_time(fabric: Optional[L.Fabric] = None):
     """
     Set seed for all processes by time.
     """
     # set seed for all processes
-    if fabric.is_global_zero:
+    if fabric is None or fabric.is_global_zero:
         seed = int(time.time())
     else:
         seed = None
-    fabric.barrier()
-    seed = fabric.broadcast(seed, src=0)
+    if fabric is not None:
+        log.debug(f"Broadcasting seed `{seed}` to all processes")
+        fabric.barrier()
+        seed = fabric.broadcast(seed, src=0)
     L.seed_everything(seed)

fusion_bench/utils/lazy_state_dict.py

@@ -1,13 +1,16 @@
 import json
 import logging
 import os
-from typing import TYPE_CHECKING, Dict, Iterator, List, Optional, Tuple
+from copy import deepcopy
+from typing import TYPE_CHECKING, Dict, Iterator, List, Optional, Tuple, Type
 
 import torch
+from accelerate import init_empty_weights
 from accelerate.utils.constants import SAFE_WEIGHTS_NAME, WEIGHTS_NAME
 from huggingface_hub import snapshot_download
 from safetensors import safe_open
 from safetensors.torch import load_file
+from torch import nn
 from transformers import AutoConfig
 
 from fusion_bench.utils.dtype import parse_dtype
@@ -59,6 +62,8 @@ class LazyStateDict:
     def __init__(
         self,
         checkpoint: str,
+        meta_module_class: Optional[Type[nn.Module]] = None,
+        meta_module: Optional[nn.Module] = None,
         cache_state_dict: bool = False,
         torch_dtype: Optional[torch.dtype] = None,
         device: str = "cpu",
@@ -66,6 +71,22 @@ class LazyStateDict:
         hf_cache_dir: Optional[str] = None,
         hf_proxies: Optional[Dict] = None,
     ):
+        self.meta_module_class = meta_module_class
+        self.meta_module = meta_module
+        if self.meta_module_class is not None:
+            if self.meta_module is not None:
+                raise ValueError(
+                    "Cannot provide both meta_module_class and meta_module, please provide only one."
+                )
+            with init_empty_weights():
+                self.meta_module = self.meta_module_class.from_pretrained(
+                    checkpoint,
+                    torch_dtype=torch_dtype,
+                    revision=hf_revision,
+                    cache_dir=hf_cache_dir,
+                    proxies=hf_proxies,
+                )
+
         self._checkpoint = checkpoint
         self._local_path = resolve_checkpoint_path(
             checkpoint,
@@ -78,10 +99,32 @@ class LazyStateDict:
             self._resolve_checkpoint_files(self._local_path)
         )
 
-        if cache_state_dict:
-            self._state_dict_cache = {}
+        if self._index is not None:
+            # if meta_module is provided, remove the keys that are not in the meta_module
+            if self.meta_module is not None:
+                meta_module_state_dict = self.meta_module.state_dict()
+                for key in tuple(self._index.keys()):
+                    if key not in meta_module_state_dict:
+                        self._index.pop(key)
+            if cache_state_dict:
+                self._state_dict_cache = {}
+            else:
+                self._state_dict_cache = None
+        elif len(self._checkpoint_files) == 1 and self._checkpoint_files[0].endswith(
+            WEIGHTS_NAME
+        ):
+            log.info(f"Loading full state dict from {WEIGHTS_NAME}")
+            self._state_dict_cache = torch.load(self._checkpoint_files[0])
+            # if meta_module is provided, remove the keys that are not in the meta_module
+            if self.meta_module is not None:
+                meta_module_state_dict = self.meta_module.state_dict()
+                for key in tuple(self._state_dict_cache.keys()):
+                    if key not in meta_module_state_dict:
+                        self._state_dict_cache.pop(key)
         else:
-            self._state_dict_cache = None
+            raise ValueError(
+                f"Cannot determine the type of checkpoint, please provide a checkpoint path to a file containing a whole state dict with file name {WEIGHTS_NAME} or {SAFE_WEIGHTS_NAME}, or the index of a sharded checkpoint ending with `.index.json`."
+            )
 
         self._torch_dtype = parse_dtype(torch_dtype)
         self._device = device
@@ -152,6 +195,8 @@ class LazyStateDict:
             checkpoint_files = [
                 os.path.join(checkpoint_folder, f) for f in checkpoint_files
             ]
+        else:
+            index = None
         return index, index_filename, checkpoint_files
 
     def _load_tensor_from_checkpoint_file(
@@ -248,16 +293,24 @@ class LazyStateDict:
     def __iter__(self) -> Iterator[str]:
         if self._index is not None:
             return iter(self._index)
-        return iter(self._checkpoint_files)
+        elif self._state_dict_cache is not None:
+            return iter(self._state_dict_cache)
+        else:
+            raise RuntimeError(
+                "Unexpected error: cannot determine the keys in the state dict."
+            )
 
-    def keys(self) -> List[str]:
-        return list(self)
+    def keys(self) -> Iterator[str]:
+        for key in self:
+            yield key
 
-    def values(self) -> List[torch.Tensor]:
-        return [self[key] for key in self]
+    def values(self) -> Iterator[torch.Tensor]:
+        for key in self:
+            yield self[key]
 
     def items(self) -> Iterator[Tuple[str, torch.Tensor]]:
-        return ((key, self[key]) for key in self)
+        for key in self:
+            yield key, self[key]
 
     def __repr__(self) -> str:
         if self._index is not None:
@@ -266,3 +319,20 @@ class LazyStateDict:
             return (
                 f"{self.__class__.__name__}(checkpoint_files={self._checkpoint_files})"
             )
+
+    def get_parameter(self, target: str) -> torch.Tensor:
+        return self[target]
+
+    def get_submodule(self, target: str) -> nn.Module:
+        if self.meta_module is not None:
+            module: nn.Module = deepcopy(self.meta_module.get_submodule(target))
+            module.to_empty(device=self._device)
+            state_dict = {}
+            for name, _ in module.named_parameters():
+                state_dict[name] = self[f"{target}.{name}"]
+            module.load_state_dict(state_dict)
+            return module
+        else:
+            raise RuntimeError(
+                "Cannot get submodule because meta_module is not provided."
+            )

fusion_bench/utils/pylogger.py

@@ -62,6 +62,8 @@ class RankZeroLogger(logging.Logger):
     def _log(self, *args, **kwargs):
         if "stacklevel" in kwargs:
             kwargs["stacklevel"] += 1
+        else:
+            kwargs["stacklevel"] = 2
         return super()._log(*args, **kwargs)
 
     def is_global_zero(self):

{fusion_bench-0.2.17.dist-info → fusion_bench-0.2.18.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: fusion_bench
-Version: 0.2.17
+Version: 0.2.18
 Summary: A Comprehensive Benchmark of Deep Model Fusion
 Author-email: Anke Tang <tang.anke@foxmail.com>
 License: MIT License