keras/qwen1.5_moe_2.7b_en

Model Overview

Model Summary

Qwen1.5-MoE is a transformer-based MoE decoder-only language model pre-trained on a large amount of data. Qwen1.5-MoE employs a Mixture of Experts (MoE) architecture, where the models are upcycled from dense language models. For instance, Qwen1.5-MoE-A2.7B is upcycled from Qwen-1.8B. It has 14.3B parameters in total and 2.7B activated parameters during runtime, while achieving comparable performance to Qwen1.5-7B, it only requires 25% of the training resources.

Weights are released under the Apache 2 License . Keras model code is released under the Apache 2 License.

Links

• Qwen 1.5 MoE Quickstart Notebook
• Qwen 1.5 MoE API Documentation
• Qwen 1.5 MoE Model Card
• KerasHub Beginner Guide
• KerasHub Model Publishing Guide

Installation

Keras and KerasHub can be installed with:

pip install -U -q keras-hub
pip install -U -q keras

Jax, TensorFlow, and Torch come preinstalled in Kaggle Notebooks. For instructions on installing them in another environment see the Keras Getting Started page.

Presets

The following model checkpoints are provided by the Keras team. Full code examples for each are available below.

Preset name	Parameters	Description
qwen1.5_moe_2.7b_en	2.7B	24-layer Qwen MoE model with 2.7 billion parameters and 8 experts per MoE layer.

Example Usage

import keras
import keras_hub
import numpy as np

# Basic text generation with Qwen MoE
qwen_moe = keras_hub.models.QwenMoeCausalLM.from_preset("qwen1.5_moe_2.7b_en")
qwen_moe.generate("I want to say", max_length=30)

# Batch generation with multiple prompts
qwen_moe.generate(["This is a", "Where are you"], max_length=30)

# Using different sampling strategies
qwen_moe = keras_hub.models.QwenMoeCausalLM.from_preset("qwen1.5_moe_2.7b_en")
# Greedy sampling
qwen_moe.compile(sampler="greedy")
qwen_moe.generate("I want to say", max_length=30)
# Beam search with MoE-specific parameters
qwen_moe.compile(
    sampler=keras_hub.samplers.BeamSampler(
        num_beams=2,
        decoder_sparse_step=2,  # MoE-specific: control expert usage frequency
        top_k_experts=2,        # MoE-specific: number of experts to use per token
    )
)
qwen_moe.generate("I want to say", max_length=30)

# Generate without preprocessing
prompt = {
    "token_ids": np.array([[15191, 374, 0, 0, 0]] * 2),
    "padding_mask": np.array([[1, 1, 0, 0, 0]] * 2),
}

qwen_moe = keras_hub.models.QwenMoeCausalLM.from_preset(
    "qwen1.5_moe_2.7b_en",
    preprocessor=None,
)
qwen_moe.generate(
    prompt,
    num_experts=8,           # Total number of experts per layer
    top_k_experts=2,         # Number of experts to use per token
    decoder_sparse_step=2,   # Control expert usage frequency
)

# Training on a single batch
features = ["The quick brown fox jumped.", "I forgot my homework."]
qwen_moe = keras_hub.models.QwenMoeCausalLM.from_preset("qwen1.5_moe_2.7b_en")
qwen_moe.fit(
    x=features,
    batch_size=2,
    router_aux_loss_coefficient=0.01,  # MoE-specific: aux loss for router training
)

# Training without preprocessing
x = {
    "token_ids": np.array([[1, 2, 3, 4, 5]] * 2),
    "padding_mask": np.array([[1, 1, 1, 1, 1]] * 2),
}
y = np.array([[2, 3, 4, 5, 0]] * 2)
sw = np.array([[1, 1, 1, 1, 1]] * 2)

qwen_moe = keras_hub.models.QwenMoeCausalLM.from_preset(
    "qwen1.5_moe_2.7b_en",
    preprocessor=None,
)
qwen_moe.fit(
    x=x,
    y=y,
    sample_weight=sw,
    batch_size=2,
    router_aux_loss_coefficient=0.01,  # MoE-specific: aux loss weight
)

Example Usage with Hugging Face URI

import keras
import keras_hub
import numpy as np

# Basic text generation with Qwen MoE
qwen_moe = keras_hub.models.QwenMoeCausalLM.from_preset("hf://keras/qwen1.5_moe_2.7b_en")
qwen_moe.generate("I want to say", max_length=30)

# Batch generation with multiple prompts
qwen_moe.generate(["This is a", "Where are you"], max_length=30)

# Using different sampling strategies
qwen_moe = keras_hub.models.QwenMoeCausalLM.from_preset("hf://keras/qwen1.5_moe_2.7b_en")
# Greedy sampling
qwen_moe.compile(sampler="greedy")
qwen_moe.generate("I want to say", max_length=30)
# Beam search with MoE-specific parameters
qwen_moe.compile(
    sampler=keras_hub.samplers.BeamSampler(
        num_beams=2,
        decoder_sparse_step=2,  # MoE-specific: control expert usage frequency
        top_k_experts=2,        # MoE-specific: number of experts to use per token
    )
)
qwen_moe.generate("I want to say", max_length=30)

# Generate without preprocessing
prompt = {
    "token_ids": np.array([[15191, 374, 0, 0, 0]] * 2),
    "padding_mask": np.array([[1, 1, 0, 0, 0]] * 2),
}

qwen_moe = keras_hub.models.QwenMoeCausalLM.from_preset(
    "hf://keras/qwen1.5_moe_2.7b_en",
    preprocessor=None,
)
qwen_moe.generate(
    prompt,
    num_experts=8,           # Total number of experts per layer
    top_k_experts=2,         # Number of experts to use per token
    decoder_sparse_step=2,   # Control expert usage frequency
)

# Training on a single batch
features = ["The quick brown fox jumped.", "I forgot my homework."]
qwen_moe = keras_hub.models.QwenMoeCausalLM.from_preset("hf://keras/qwen1.5_moe_2.7b_en")
qwen_moe.fit(
    x=features,
    batch_size=2,
    router_aux_loss_coefficient=0.01,  # MoE-specific: aux loss for router training
)

# Training without preprocessing
x = {
    "token_ids": np.array([[1, 2, 3, 4, 5]] * 2),
    "padding_mask": np.array([[1, 1, 1, 1, 1]] * 2),
}
y = np.array([[2, 3, 4, 5, 0]] * 2)
sw = np.array([[1, 1, 1, 1, 1]] * 2)

qwen_moe = keras_hub.models.QwenMoeCausalLM.from_preset(
    "hf://keras/qwen1.5_moe_2.7b_en",
    preprocessor=None,
)
qwen_moe.fit(
    x=x,
    y=y,
    sample_weight=sw,
    batch_size=2,
    router_aux_loss_coefficient=0.01,  # MoE-specific: aux loss weight
)