fusion-bench 0.2.13__tar.gz → 0.2.15__tar.gz

{fusion_bench-0.2.13 → fusion_bench-0.2.15}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: fusion_bench
-Version: 0.2.13
+Version: 0.2.15
 Summary: A Comprehensive Benchmark of Deep Model Fusion
 Author-email: Anke Tang <tang.anke@foxmail.com>
 License: MIT License
@@ -45,6 +45,8 @@ Requires-Dist: rich
 Requires-Dist: scipy
 Requires-Dist: h5py
 Requires-Dist: pytest
+Provides-Extra: lm-eval-harness
+Requires-Dist: lm-eval; extra == "lm-eval-harness"
 Dynamic: license-file
 
 <div align='center'>
@@ -61,7 +63,7 @@ Dynamic: license-file
 
 </div>
 
-> [!TIP]  
+> [!TIP]
 > Documentation is available at [tanganke.github.io/fusion_bench/](https://tanganke.github.io/fusion_bench/).
 
 ## Overview
@@ -122,7 +124,7 @@ Merging multiple expert models offers a promising approach for performing multi-
 
 ## Installation
 
-install from PyPI:
+Install from PyPI:
 
 ```bash
 pip install fusion-bench
@@ -137,6 +139,27 @@ cd fusion_bench
 pip install -e . # install the package in editable mode
 ```
 
+### Install with [Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness)
+
+[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.10256836.svg)](https://doi.org/10.5281/zenodo.10256836)
+
+
+```bash
+pip install "fusion-bench[lm-eval-harness]"
+```
+
+or install from local directory
+
+```bash
+pip install -e ".[lm-eval-harness]"
+```
+
+This will install the latest version of fusion-bench and the dependencies required for LM-Eval Harness.
+Documentation for using LM-Eval Harness within FusionBench framework can be found at [this online documentation](https://tanganke.github.io/fusion_bench/taskpool/lm_eval_harness) or in the [`docs/taskpool/lm_eval_harness.md`](docs/taskpool/lm_eval_harness.md) markdown file.
+
+> [!TIP]
+> Documentation for merging large language models using FusionBench can be found at [this online documentation](https://tanganke.github.io/fusion_bench/modelpool/causal_lm) or in the [`docs/modelpool/causal_lm.md`](docs/modelpool/causal_lm.md) markdown file.
+
 ## Introduction to Deep Model Fusion
 
 Deep model fusion is a technique that merges, ensemble, or fuse multiple deep neural networks to obtain a unified model.

{fusion_bench-0.2.13 → fusion_bench-0.2.15}/README.md

@@ -12,7 +12,7 @@
 
 </div>
 
-> [!TIP]  
+> [!TIP]
 > Documentation is available at [tanganke.github.io/fusion_bench/](https://tanganke.github.io/fusion_bench/).
 
 ## Overview
@@ -73,7 +73,7 @@ Merging multiple expert models offers a promising approach for performing multi-
 
 ## Installation
 
-install from PyPI:
+Install from PyPI:
 
 ```bash
 pip install fusion-bench
@@ -88,6 +88,27 @@ cd fusion_bench
 pip install -e . # install the package in editable mode
 ```
 
+### Install with [Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness)
+
+[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.10256836.svg)](https://doi.org/10.5281/zenodo.10256836)
+
+
+```bash
+pip install "fusion-bench[lm-eval-harness]"
+```
+
+or install from local directory
+
+```bash
+pip install -e ".[lm-eval-harness]"
+```
+
+This will install the latest version of fusion-bench and the dependencies required for LM-Eval Harness.
+Documentation for using LM-Eval Harness within FusionBench framework can be found at [this online documentation](https://tanganke.github.io/fusion_bench/taskpool/lm_eval_harness) or in the [`docs/taskpool/lm_eval_harness.md`](docs/taskpool/lm_eval_harness.md) markdown file.
+
+> [!TIP]
+> Documentation for merging large language models using FusionBench can be found at [this online documentation](https://tanganke.github.io/fusion_bench/modelpool/causal_lm) or in the [`docs/modelpool/causal_lm.md`](docs/modelpool/causal_lm.md) markdown file.
+
 ## Introduction to Deep Model Fusion
 
 Deep model fusion is a technique that merges, ensemble, or fuse multiple deep neural networks to obtain a unified model.

{fusion_bench-0.2.13 → fusion_bench-0.2.15}/fusion_bench/method/ada_svd/clip_vision.py

@@ -31,7 +31,10 @@ from fusion_bench.method import WeightedAverageAlgorithm
 from fusion_bench.method.simple_average import simple_average
 from fusion_bench.mixins import SimpleProfilerMixin
 from fusion_bench.modelpool import CLIPVisionModelPool
-from fusion_bench.models.smile_moe.linear import ExpertNotTrainedError, SmileMoELinear
+from fusion_bench.models.smile_moe.linear_from_module import (
+    ExpertNotTrainedError,
+    SmileMoELinear,
+)
 from fusion_bench.models.utils import find_layers_with_type, get_attr, set_attr
 from fusion_bench.utils.devices import get_device
 

{fusion_bench-0.2.13 → fusion_bench-0.2.15}/fusion_bench/method/smile_upscaling/smile_mistral_upscaling.py

@@ -9,11 +9,16 @@ import torch.nn.functional as F
 from accelerate import init_empty_weights
 from torch import Tensor, nn
 from tqdm.auto import tqdm
-from transformers import AutoConfig, AutoTokenizer, MistralForCausalLM
+from transformers import (
+    AutoConfig,
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    MistralForCausalLM,
+)
 from transformers.models.mistral.modeling_mistral import MistralDecoderLayer
 
-from fusion_bench.compat.method import ModelFusionAlgorithm
 from fusion_bench.compat.modelpool import to_modelpool
+from fusion_bench.method import BaseAlgorithm
 from fusion_bench.method.simple_average import simple_average
 from fusion_bench.mixins.simple_profiler import SimpleProfilerMixin
 from fusion_bench.modelpool import BaseModelPool
@@ -25,158 +30,23 @@ from fusion_bench.models.modeling_smile_mistral.modeling_smile_mistral import (
     SmileLinear,
     SmileMistralDecoderLayer,
 )
-from fusion_bench.models.utils import get_attr, set_attr
+from fusion_bench.models.smile_moe.linear_from_hf_config import (
+    ExpertNotTrainedError,
+    upscale_to_smile_linear,
+)
 from fusion_bench.utils.dtype import parse_dtype
 from fusion_bench.utils.parameters import print_parameters
-from fusion_bench.utils.state_dict_arithmetic import state_dict_sub
 
 log = logging.getLogger(__name__)
 
 
-class ExpertNotTrainedError(Exception):
-    pass
-
-
-def _is_all_zeros(tensor: Tensor | List[Tensor]) -> bool:
-    """
-    Check if a tensor or a list of tensors are all zeros.
-
-    Args:
-        tensor (Tensor | List[Tensor]): The tensor or list of tensors to check.
-
-    Returns:
-        bool: True if all elements are zeros, False otherwise.
-    """
-    if isinstance(tensor, Tensor):
-        return torch.allclose(tensor, torch.zeros_like(tensor))
-    else:
-        return all(_is_all_zeros(t) for t in tensor)
-
-
-def _svd(w: Tensor, full_matrices=False) -> Tuple[Tensor, Tensor, Tensor]:
-    """
-    Perform Singular Value Decomposition (SVD) on a tensor.
-
-    Args:
-        w (Tensor): The input tensor.
-        full_matrices (bool, optional): Whether to compute the full-sized U and V matrices. Defaults to False.
-
-    Returns:
-        Tuple[Tensor, Tensor, Tensor]: The U, S, and V matrices from SVD.
-    """
-    device = w.device
-    if w.device != torch.float32 or w.device != torch.float64:
-        w = w.float()
-
-    u, s, vh = torch.linalg.svd(
-        w,
-        full_matrices=full_matrices,
-        # driver="gesvd" if w.is_cuda else None
-    )
-    v = vh.T
-
-    u = u.to(device)
-    s = s.to(device)
-    v = v.to(device)
-    return u, s, v
-
-
-def svd(
-    w: Tensor, full_matrices=True, accelerator=None
-) -> Tuple[Tensor, Tensor, Tensor]:
-    """
-    Perform SVD on a tensor with optional acceleration.
-
-    Args:
-        w (Tensor): The input tensor.
-        full_matrices (bool, optional): Whether to compute the full-sized U and V matrices. Defaults to True.
-        accelerator (optional): The device to perform the computation on. Defaults to None.
-
-    Returns:
-        Tuple[Tensor, Tensor, Tensor]: The U, S, and V matrices from SVD.
-    """
-    if accelerator is None:
-        return _svd(w, full_matrices=full_matrices)
-    original_device = w.device
-    w = w.to(accelerator)
-    u, s, v = _svd(w)
-    return u, s, v
-
-
-@torch.no_grad()
-def upscale_to_smile_linear(
-    base: nn.Linear, experts: List[nn.Linear], target: SmileLinear, accelerator=None
-):
-    """
-    Upscale a base linear layer to a SmileLinear layer using expert models.
-
-    Args:
-        base (nn.Linear): The base linear layer.
-        experts (List[nn.Linear]): A list of expert linear layers.
-        target (SmileLinear): The target SmileLinear layer.
-        accelerator (optional): The device to perform the computation on. Defaults to None.
-
-    Returns:
-        SmileLinear: The upscaled SmileLinear layer.
-    """
-    w = base.weight
-    w_ft_list = [e.weight for e in experts]
-    dw_list = [w_ft - w for w_ft in w_ft_list]
-
-    if _is_all_zeros(dw_list):
-        raise ExpertNotTrainedError("Expert models are not trained")
-
-    rank_of_router = target.rank_of_router
-    rank_of_expert = target.rank_of_expert
-    num_local_experts = target.num_local_experts
-    svd_list = [svd(dw, accelerator=accelerator) for dw in dw_list]
-
-    # gate
-    gate_weight = []
-    for u, s, v in svd_list:
-        gate_weight.append(v[:, :rank_of_router].T)
-    gate_weight = (
-        torch.stack(gate_weight, dim=0)
-        .reshape(num_local_experts * rank_of_router, -1)
-        .contiguous()
-    )
-
-    target.gate.load_state_dict({"weight": gate_weight})
-
-    # shared linear
-    target.shared_linear.load_state_dict(base.state_dict())
-
-    # experts
-    if rank_of_expert > 0:
-        for expert_idx, target_expert in enumerate(target.experts):
-            u, s, v = svd_list[expert_idx]
-            u = u[:, :rank_of_expert]
-            s = s[:rank_of_expert]
-            v = v[:, :rank_of_expert]
-            state_dict = {"u": u, "svh": (s * v).T}
-            if experts[expert_idx].bias is not None:
-                state_dict["bias"] = experts[expert_idx].bias.data
-            target_expert.load_state_dict(state_dict)
-    else:
-        for expert_idx, target_expert in enumerate(target.experts):
-            target_expert.load_state_dict(
-                state_dict_sub(experts[expert_idx].state_dict(), base.state_dict())
-            )
-
-    return target
-
-
-class SmileMistralUpscalingAlgorithm(ModelFusionAlgorithm, SimpleProfilerMixin):
+class SmileMistralUpscalingAlgorithm(BaseAlgorithm, SimpleProfilerMixin):
     R"""
     SmileMistralUpscalingAlgorithm is a model fusion algorithm designed to upscale
     a pretrained Mistral model using a set of fine-tuned expert models. The algorithm
     leverages Singular Value Decomposition (SVD) to merge the weights of the pretrained
     model and the expert models into a new upscaled model.
 
-    Attributes:
-        modelpool (BaseModelPool): The pool of models to be used for upscaling.
-        config (dict): Configuration parameters for the upscaling process.
-
     Methods:
         run(modelpool: BaseModelPool) -> SmileMistralForCausalLM:
             Executes the upscaling process and returns the upscaled model.
@@ -185,6 +55,37 @@ class SmileMistralUpscalingAlgorithm(ModelFusionAlgorithm, SimpleProfilerMixin):
             Merges the pretrained model with the fine-tuned models to create an upscaled model.
     """
 
+    _config_mapping = BaseAlgorithm._config_mapping | {
+        "device": "device",
+        "accelerator": "accelerator",
+        "model_path": "model_path",
+        "model_dtype": "model_dtype",
+        "num_experts_per_tok": "num_experts_per_tok",
+        "rank_of_router": "rank_of_router",
+        "rank_of_expert": "rank_of_expert",
+    }
+
+    def __init__(
+        self,
+        device,
+        accelerator,
+        model_path,
+        model_dtype,
+        num_experts_per_tok,
+        rank_of_router,
+        rank_of_expert,
+        **kwargs,
+    ):
+        self.device = device
+        self.accelerator = accelerator
+        self.model_path = model_path
+        self.model_dtype = model_dtype
+        # SmileMoE parameters, except `num_local_experts` which is set later according to the number of finetuned models
+        self.num_experts_per_tok = num_experts_per_tok
+        self.rank_of_router = rank_of_router
+        self.rank_of_expert = rank_of_expert
+        super().__init__(**kwargs)
+
     @torch.no_grad()
     def run(self, modelpool: BaseModelPool) -> SmileMistralForCausalLM:
         """
@@ -199,15 +100,15 @@ class SmileMistralUpscalingAlgorithm(ModelFusionAlgorithm, SimpleProfilerMixin):
         self.modelpool = modelpool = to_modelpool(modelpool)
         config = self.config
 
-        print(config)
+        # load model from path if provided and return directly
         if config.model_path is not None and os.path.exists(config.model_path):
             log.info(f"Loading model from {config.model_path}")
-            model = torch.load(config.model_path)
+            model = AutoModelForCausalLM.from_pretrained(config.model_path)
             print_parameters(model)
             return model
 
         with self.profile("load pretrained model"):
-            pretrained_model = modelpool.load_model("_pretrained_")
+            pretrained_model = modelpool.load_pretrained_model()
         with self.profile("load fine-tuned model"):
             finetuned_models = [
                 m for m in tqdm(modelpool.models(), total=len(modelpool.model_names))

fusion_bench-0.2.15/fusion_bench/method/smile_upscaling/smile_qwen2_upscaling.py

@@ -0,0 +1,229 @@
+import logging
+import os
+from copy import deepcopy
+from typing import TYPE_CHECKING, Dict, List, Tuple
+
+import torch
+from accelerate import init_empty_weights
+from tqdm.auto import tqdm
+from transformers import (
+    AutoConfig,
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    Qwen2ForCausalLM,
+)
+from transformers.models.qwen2.modeling_qwen2 import Qwen2DecoderLayer
+
+from fusion_bench import BaseAlgorithm, BaseModelPool
+from fusion_bench.compat.modelpool import to_modelpool
+from fusion_bench.mixins import SimpleProfilerMixin
+from fusion_bench.models.modeling_smile_qwen2 import (
+    SmileQwen2Config,
+    SmileQwen2ForCausalLM,
+)
+from fusion_bench.models.modeling_smile_qwen2.modeling_smile_qwen2 import (
+    SmileQwen2DecoderLayer,
+)
+from fusion_bench.models.smile_moe.linear_from_hf_config import (
+    ExpertNotTrainedError,
+    upscale_to_smile_linear,
+)
+from fusion_bench.utils.dtype import parse_dtype
+from fusion_bench.utils.parameters import print_parameters
+
+log = logging.getLogger(__name__)
+
+
+class SmileQwen2UpscalingAlgorithm(BaseAlgorithm, SimpleProfilerMixin):
+    R"""
+    SmileQwen2UpscalingAlgorithm is a model fusion algorithm designed to upscale
+    a pretrained Qwen2 model using a set of fine-tuned expert models. The algorithm
+    leverages Singular Value Decomposition (SVD) to merge the weights of the pretrained
+    model and the expert models into a new upscaled model.
+
+    Methods:
+        run(modelpool: BaseModelPool) -> SmileQwen2ForCausalLM:
+            Executes the upscaling process and returns the upscaled model.
+
+        merge(pretrained_model: Qwen2ForCausalLM, finetuned_models: List[Qwen2ForCausalLM]) -> SmileQwen2ForCausalLM:
+            Merges the pretrained model with the fine-tuned models to create an upscaled model.
+    """
+
+    _config_mapping = BaseAlgorithm._config_mapping | {
+        "device": "device",
+        "accelerator": "accelerator",
+        "model_path": "model_path",
+        "model_dtype": "model_dtype",
+        "num_experts_per_tok": "num_experts_per_tok",
+        "rank_of_router": "rank_of_router",
+        "rank_of_expert": "rank_of_expert",
+    }
+
+    def __init__(
+        self,
+        device,
+        accelerator,
+        model_path,
+        model_dtype,
+        num_experts_per_tok,
+        rank_of_router,
+        rank_of_expert,
+        **kwargs,
+    ):
+        self.device = device
+        self.accelerator = accelerator
+        self.model_path = model_path
+        self.model_dtype = model_dtype
+        # SmileMoE parameters, except `num_local_experts` which is set later according to the number of finetuned models
+        self.num_experts_per_tok = num_experts_per_tok
+        self.rank_of_router = rank_of_router
+        self.rank_of_expert = rank_of_expert
+        super().__init__(**kwargs)
+
+    @torch.no_grad()
+    def run(self, modelpool: BaseModelPool) -> SmileQwen2ForCausalLM:
+        """
+        Executes the upscaling process.
+
+        Args:
+            modelpool (ModelPool): The pool of models to be used for upscaling.
+
+        Returns:
+            SmileQwen2ForCausalLM: The upscaled model.
+        """
+        self.modelpool = modelpool = to_modelpool(modelpool)
+        config = self.config
+
+        # load model from path if provided and return directly
+        if config.model_path is not None and os.path.exists(config.model_path):
+            log.info(f"Loading model from {config.model_path}")
+            model = AutoModelForCausalLM.from_pretrained(config.model_path)
+            print_parameters(model)
+            return model
+
+        with self.profile("load pretrained model"):
+            pretrained_model = modelpool.load_pretrained_model()
+        with self.profile("load fine-tuned model"):
+            finetuned_models = [
+                m for m in tqdm(modelpool.models(), total=len(modelpool.model_names))
+            ]
+
+        if config.device == "cuda" and torch.cuda.is_available():
+            pretrained_model = pretrained_model.cuda()
+            print("parameter count of pretrained model:")
+            print_parameters(pretrained_model)
+            finetuned_models = [m.cuda() for m in finetuned_models]
+
+        with self.profile("merge model"):
+            model = self.merge(pretrained_model, finetuned_models)
+
+        self.print_profile_summary()
+        print("parameter count of upscaled MoE model:")
+        print_parameters(model)
+        print(model)
+
+        if config.model_dtype is not None:
+            model.to(dtype=parse_dtype(config.model_dtype))
+
+        if config.model_path is not None:
+            if os.path.dirname(config.model_path):
+                os.makedirs(os.path.dirname(config.model_path), exist_ok=True)
+            log.info(f"Saving model to {config.model_path}")
+            pretrained_model_config = self.modelpool.get_model_config("_pretrained_")
+            pretrained_path = pretrained_model_config.get(
+                "path", pretrained_model_config["pretrained_model_name_or_path"]
+            )
+            tokenizer = AutoTokenizer.from_pretrained(pretrained_path)
+            tokenizer.save_pretrained(config.model_path)
+            model.save_pretrained(config.model_path)
+
+        return model
+
+    def merge(
+        self,
+        pretrained_model: Qwen2ForCausalLM,
+        finetuned_models: List[Qwen2ForCausalLM],
+    ):
+        """
+        Merges the pretrained model with the fine-tuned models to create an upscaled model.
+
+        Args:
+            pretrained_model (Qwen2ForCausalLM): The pretrained model.
+            finetuned_models (List[Qwen2ForCausalLM]): A list of fine-tuned models.
+
+        Returns:
+            SmileQwen2ForCausalLM: The upscaled model.
+        """
+        config = self.config
+
+        with init_empty_weights():
+            pretrained_model_config = self.modelpool.get_model_config("_pretrained_")
+            pretrained_path = pretrained_model_config.get(
+                "path", pretrained_model_config["pretrained_model_name_or_path"]
+            )
+            base_config = AutoConfig.from_pretrained(pretrained_path)
+            model_config = SmileQwen2Config(
+                num_experts_per_tok=config.num_experts_per_tok,
+                rank_of_router=config.rank_of_router,
+                rank_of_expert=config.rank_of_expert,
+                num_local_experts=len(finetuned_models),
+                **base_config.to_dict(),
+            )
+            model = SmileQwen2ForCausalLM(model_config)
+
+        model.to(dtype=pretrained_model.dtype).to_empty(device="cpu")
+
+        # copy pretrained model weights
+        state_dict = model.state_dict()
+        pretrained_state_dict = dict(pretrained_model.state_dict())
+        for key in list(pretrained_state_dict.keys()):
+            if key not in state_dict:
+                pretrained_state_dict.pop(key)
+        model.load_state_dict(pretrained_state_dict, strict=False)
+
+        # upscale model
+        for layer_idx in tqdm(
+            range(len(pretrained_model.model.layers)),
+            "Upscaling Modules (layer)",
+            dynamic_ncols=True,
+        ):
+            pretrained_layer: Qwen2DecoderLayer = pretrained_model.model.layers[
+                layer_idx
+            ]
+            finetuned_layers: List[Qwen2DecoderLayer] = [
+                m.model.layers[layer_idx] for m in finetuned_models
+            ]
+
+            target_layer: SmileQwen2DecoderLayer = model.model.layers[layer_idx]
+
+            for n in ["q_proj", "k_proj", "v_proj", "o_proj"]:
+                try:
+                    upscale_to_smile_linear(
+                        base=getattr(pretrained_layer.self_attn, n),
+                        experts=[getattr(m.self_attn, n) for m in finetuned_layers],
+                        target=getattr(target_layer.self_attn, n),
+                        accelerator=config.accelerator,
+                    )
+                except ExpertNotTrainedError:
+                    setattr(
+                        target_layer.self_attn,
+                        n,
+                        getattr(pretrained_layer.self_attn, n),
+                    )
+
+            for n in ["gate_proj", "up_proj", "down_proj"]:
+                try:
+                    upscale_to_smile_linear(
+                        base=getattr(pretrained_layer.mlp, n),
+                        experts=[getattr(m.mlp, n) for m in finetuned_layers],
+                        target=getattr(target_layer.mlp, n),
+                        accelerator=config.accelerator,
+                    )
+                except ExpertNotTrainedError:
+                    setattr(
+                        target_layer.mlp,
+                        n,
+                        getattr(pretrained_layer.mlp, n),
+                    )
+
+        return model