Hugging Face's logo

gabrielmbmb
/
Upcycled-Qwen1.5-MoE2.7B

Text Generation
Transformers
Safetensors
qwen2_moe
conversational
Inference Endpoints
Model card Files Community
Upcycled-Qwen1.5-MoE2.7B / README.md
gabrielmbmb's picture
gabrielmbmb HF staff
Update README.md
acdd089 verified
preview code
|
raw
history blame contribute delete
9.88 kB
metadata
library_name: transformers
tags: []

Model Card for Model ID

Model Details

This is my attemp (probably too naive) to reproduce the upcycling process used to initialize Qwen1.5-MoE-A2.7B using Qwen1.5-1.8B.

Upcycling script

Script: 
from torch import nn
from transformers import AutoModelForCausalLM
from dataclasses import dataclass
from transformers import AutoModel
from typing_extensions import Self
from copy import deepcopy


@dataclass
class UpcyclingConfig:
    finegrained_experts: int
    partitions_from_mlp: int

    @property
    def upcycling_factor(self) -> int:
        return self.finegrained_experts // self.partitions_from_mlp


def iterate_in_chunks(list1, list2, chunk_size1, chunk_size2):
    iterations = max(len(list1) // chunk_size1, len(list2) // chunk_size2)
    for i in range(iterations):
        start_idx1 = i * chunk_size1
        end_idx1 = start_idx1 + chunk_size1
        start_idx2 = i * chunk_size2
        end_idx2 = start_idx2 + chunk_size2
        yield (list1[start_idx1:end_idx1], list2[start_idx2:end_idx2])


def chunk_linear(linear: nn.Linear, chunks: int, down_proj: bool = False) -> tuple[nn.Linear, ...]:
    if not down_proj:
        in_features = linear.out_features // chunks
        out_features = linear.in_features 
        dim = 0
    else:
        in_features = linear.out_features
        out_features = linear.in_features // chunks
        dim = 1

    weight = linear.weight.reshape(linear.out_features, linear.in_features)
    weights = weight.chunk(chunks, dim=dim)
    biases = linear.bias.chunk(chunks) if linear.bias is not None else [None] * chunks
    linear_layers = []
    for weight, bias in zip(weights, biases):
        new_linear = nn.Linear(
            in_features=in_features, out_features=out_features, bias=bias is not None
        )
        new_linear.weight = nn.Parameter(weight.clone())  # Clone weights to ensure they are not shared
        if bias is not None:
            new_linear.bias = nn.Parameter(bias.clone())  # Clone bias if it exists
        linear_layers.append(new_linear)
    return tuple(linear_layers)


class UpcycledModelMixin:
    sparse_moe_block_cls: type

    @classmethod
    def upcycled_from(cls, source_model, config: UpcyclingConfig) -> Self:
        upcycled_model_config = cls.config_class(**source_model.config.to_dict())
        upcycled_model_config.moe_intermediate_size = upcycled_model_config.intermediate_size // config.partitions_from_mlp
        if hasattr(upcycled_model_config, "shared_expert_intermediate_size"):
            upcycled_model_config.shared_expert_intermediate_size = source_model.config.intermediate_size

        upcycled_model = cls(upcycled_model_config)
        upcycled_model.model.embed_tokens = source_model.model.embed_tokens

        for upcycled_layer, layer in zip(upcycled_model.model.layers, source_model.model.layers):
            upcycled_layer.self_attn = layer.self_attn
            upcycled_mlp_layers = [deepcopy(layer.mlp) for _ in range(config.upcycling_factor)]

            if hasattr(upcycled_layer.mlp, "shared_expert"):
                upcycled_layer.mlp.shared_expert = upcycled_mlp_layers.pop(-1)

            for experts, mlp in iterate_in_chunks(upcycled_layer.mlp.experts, upcycled_mlp_layers, 4, 1):
                gate_projs = chunk_linear(mlp[0].gate_proj, 4, down_proj=False)
                up_projs = chunk_linear(mlp[0].up_proj, 4, down_proj=False)
                down_projs = chunk_linear(mlp[0].down_proj, 4, down_proj=True)
                for i, expert in enumerate(experts):
                    expert.gate_proj = gate_projs[i]
                    expert.up_proj = up_projs[i]
                    expert.down_proj = down_projs[i]
                    expert.act_fn = deepcopy(mlp[0].act_fn)

            upcycled_layer.input_layernorm = layer.input_layernorm
            upcycled_layer.post_attention_layernorm = layer.post_attention_layernorm

        upcycled_model.lm_head = source_model.lm_head
        return upcycled_model


from transformers import Qwen2MoeForCausalLM as _Qwen2MoeForCausalLM
from transformers.models.qwen2.modeling_qwen2 import Qwen2MLP
from transformers.models.qwen2_moe.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock

class Qwen2MoeForCausalLM(UpcycledModelMixin, _Qwen2MoeForCausalLM):
    sparse_moe_block_cls = Qwen2MoeSparseMoeBlock


source_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen1.5-1.8B", device_map="auto")
model = Qwen2MoeForCausalLM.upcycled_from(
    source_model,
    UpcyclingConfig(
        finegrained_experts=64,
        partitions_from_mlp=4,
    ),
)

model = model.bloat16()

Model Description

This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.

  • Developed by: [More Information Needed]
  • Funded by [optional]: [More Information Needed]
  • Shared by [optional]: [More Information Needed]
  • Model type: [More Information Needed]
  • Language(s) (NLP): [More Information Needed]
  • License: [More Information Needed]
  • Finetuned from model [optional]: [More Information Needed]

Model Sources [optional]

  • Repository: [More Information Needed]
  • Paper [optional]: [More Information Needed]
  • Demo [optional]: [More Information Needed]

Uses

Direct Use

[More Information Needed]

Downstream Use [optional]

[More Information Needed]

Out-of-Scope Use

[More Information Needed]

Bias, Risks, and Limitations

[More Information Needed]

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.

How to Get Started with the Model

Use the code below to get started with the model.

[More Information Needed]

Training Details

Training Data

[More Information Needed]

Training Procedure

Preprocessing [optional]

[More Information Needed]

Training Hyperparameters

  • Training regime: [More Information Needed]

Speeds, Sizes, Times [optional]

[More Information Needed]

Evaluation

Testing Data, Factors & Metrics

Testing Data

[More Information Needed]

Factors

[More Information Needed]

Metrics

[More Information Needed]

Results

[More Information Needed]

Summary

Model Examination [optional]

[More Information Needed]

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

  • Hardware Type: [More Information Needed]
  • Hours used: [More Information Needed]
  • Cloud Provider: [More Information Needed]
  • Compute Region: [More Information Needed]
  • Carbon Emitted: [More Information Needed]

Technical Specifications [optional]

Model Architecture and Objective

[More Information Needed]

Compute Infrastructure

[More Information Needed]

Hardware

[More Information Needed]

Software

[More Information Needed]

Citation [optional]

BibTeX:

[More Information Needed]

APA:

[More Information Needed]

Glossary [optional]

[More Information Needed]

More Information [optional]

[More Information Needed]

Model Card Authors [optional]

[More Information Needed]

Model Card Contact

[More Information Needed]