transformers/tests/models/qwen3_moe/test_modeling_qwen3_moe.py

# Copyright 2024 The Qwen team, Alibaba Group and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Testing suite for the PyTorch Qwen3MoE model."""

import gc
import unittest

import pytest

from transformers import AutoTokenizer, Qwen3MoeConfig, is_torch_available, set_seed
from transformers.testing_utils import (
    backend_empty_cache,
    require_bitsandbytes,
    require_flash_attn,
    require_torch,
    require_torch_gpu,
    require_torch_sdpa,
    slow,
    torch_device,
)

from ...generation.test_utils import GenerationTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin


if is_torch_available():
    import torch

    from transformers import (
        Qwen3MoeForCausalLM,
        Qwen3MoeForQuestionAnswering,
        Qwen3MoeForSequenceClassification,
        Qwen3MoeForTokenClassification,
        Qwen3MoeModel,
    )


class Qwen3MoeModelTester:
    def __init__(
        self,
        parent,
        batch_size=13,
        seq_length=7,
        is_training=True,
        use_input_mask=True,
        use_token_type_ids=True,
        use_labels=True,
        vocab_size=99,
        hidden_size=64,
        num_hidden_layers=5,
        max_window_layers=3,
        use_sliding_window=True,
        sliding_window=50,
        num_attention_heads=4,
        num_key_value_heads=2,
        head_dim=16,
        intermediate_size=37,
        hidden_act="gelu",
        hidden_dropout_prob=0.1,
        attention_probs_dropout_prob=0.1,
        max_position_embeddings=512,
        expert_interval=1,
        moe_intermediate_size=12,
        num_experts_per_tok=2,
        num_experts=8,
        norm_topk_prob=False,
        output_router_logits=False,
        router_aux_loss_coef=0.001,
        type_vocab_size=16,
        type_sequence_label_size=2,
        initializer_range=0.02,
        num_labels=3,
        num_choices=4,
        pad_token_id=0,
        bos_token_id=1,
        scope=None,
    ):
        self.parent = parent
        self.batch_size = batch_size
        self.seq_length = seq_length
        self.is_training = is_training
        self.use_input_mask = use_input_mask
        self.use_token_type_ids = use_token_type_ids
        self.use_labels = use_labels
        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.num_hidden_layers = num_hidden_layers
        self.max_window_layers = max_window_layers
        self.use_sliding_window = use_sliding_window
        self.sliding_window = sliding_window
        self.num_attention_heads = num_attention_heads
        self.num_key_value_heads = num_key_value_heads
        self.head_dim = head_dim
        self.intermediate_size = intermediate_size
        self.hidden_act = hidden_act
        self.hidden_dropout_prob = hidden_dropout_prob
        self.attention_probs_dropout_prob = attention_probs_dropout_prob
        self.max_position_embeddings = max_position_embeddings
        self.type_vocab_size = type_vocab_size
        self.type_sequence_label_size = type_sequence_label_size
        self.initializer_range = initializer_range
        self.num_labels = num_labels
        self.num_choices = num_choices
        self.pad_token_id = pad_token_id
        self.bos_token_id = bos_token_id
        self.scope = scope
        self.expert_interval = expert_interval
        self.moe_intermediate_size = moe_intermediate_size
        self.num_experts_per_tok = num_experts_per_tok
        self.num_experts = num_experts
        self.norm_topk_prob = norm_topk_prob
        self.output_router_logits = output_router_logits
        self.router_aux_loss_coef = router_aux_loss_coef

    # Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs
    def prepare_config_and_inputs(self):
        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)

        input_mask = None
        if self.use_input_mask:
            input_mask = torch.tril(torch.ones_like(input_ids).to(torch_device))

        token_type_ids = None
        if self.use_token_type_ids:
            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)

        sequence_labels = None
        token_labels = None
        choice_labels = None
        if self.use_labels:
            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
            choice_labels = ids_tensor([self.batch_size], self.num_choices)

        config = self.get_config()

        return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels

    def get_config(self):
        return Qwen3MoeConfig(
            vocab_size=self.vocab_size,
            hidden_size=self.hidden_size,
            num_hidden_layers=self.num_hidden_layers,
            max_window_layers=self.max_window_layers,
            use_sliding_window=self.use_sliding_window,
            sliding_window=self.sliding_window,
            num_attention_heads=self.num_attention_heads,
            num_key_value_heads=self.num_key_value_heads,
            head_dim=self.head_dim,
            intermediate_size=self.intermediate_size,
            hidden_act=self.hidden_act,
            hidden_dropout_prob=self.hidden_dropout_prob,
            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
            max_position_embeddings=self.max_position_embeddings,
            expert_interval=self.expert_interval,
            moe_intermediate_size=self.moe_intermediate_size,
            num_experts_per_tok=self.num_experts_per_tok,
            num_experts=self.num_experts,
            norm_topk_prob=self.norm_topk_prob,
            output_router_logits=self.output_router_logits,
            router_aux_loss_coef=self.router_aux_loss_coef,
            type_vocab_size=self.type_vocab_size,
            is_decoder=False,
            initializer_range=self.initializer_range,
            pad_token_id=self.pad_token_id,
            bos_token_id=self.bos_token_id,
        )

    # Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model with Llama->Qwen3Moe
    def create_and_check_model(
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = Qwen3MoeModel(config=config)
        model.to(torch_device)
        model.eval()
        result = model(input_ids, attention_mask=input_mask)
        result = model(input_ids)
        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))

    # Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common
    def prepare_config_and_inputs_for_common(self):
        config_and_inputs = self.prepare_config_and_inputs()
        (
            config,
            input_ids,
            token_type_ids,
            input_mask,
            sequence_labels,
            token_labels,
            choice_labels,
        ) = config_and_inputs
        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
        return config, inputs_dict


@require_torch
# Copied from tests.models.mistral.test_modeling_mistral.MistralModelTest with Mistral->Qwen3Moe
class Qwen3MoeModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
    all_model_classes = (
        (
            Qwen3MoeModel,
            Qwen3MoeForCausalLM,
            Qwen3MoeForSequenceClassification,
            Qwen3MoeForTokenClassification,
            Qwen3MoeForQuestionAnswering,
        )
        if is_torch_available()
        else ()
    )
    pipeline_model_mapping = (
        {
            "feature-extraction": Qwen3MoeModel,
            "text-classification": Qwen3MoeForSequenceClassification,
            "token-classification": Qwen3MoeForTokenClassification,
            "text-generation": Qwen3MoeForCausalLM,
            "zero-shot": Qwen3MoeForSequenceClassification,
            "question-answering": Qwen3MoeForQuestionAnswering,
        }
        if is_torch_available()
        else {}
    )
    test_headmasking = False
    test_pruning = False
    fx_compatible = False  # Broken by attention refactor cc @Cyrilvallez

    # TODO (ydshieh): Check this. See https://app.circleci.com/pipelines/github/huggingface/transformers/79245/workflows/9490ef58-79c2-410d-8f51-e3495156cf9c/jobs/1012146
    def is_pipeline_test_to_skip(
        self,
        pipeline_test_case_name,
        config_class,
        model_architecture,
        tokenizer_name,
        image_processor_name,
        feature_extractor_name,
        processor_name,
    ):
        return True

    def setUp(self):
        self.model_tester = Qwen3MoeModelTester(self)
        self.config_tester = ConfigTester(self, config_class=Qwen3MoeConfig, hidden_size=37)

    def test_config(self):
        self.config_tester.run_common_tests()

    def test_model(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_model(*config_and_inputs)

    def test_model_various_embeddings(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        for type in ["absolute", "relative_key", "relative_key_query"]:
            config_and_inputs[0].position_embedding_type = type
            self.model_tester.create_and_check_model(*config_and_inputs)

    def test_torch_fx_output_loss(self):
        super().test_torch_fx_output_loss()

    def test_Qwen3Moe_sequence_classification_model(self):
        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.num_labels = 3
        input_ids = input_dict["input_ids"]
        attention_mask = input_ids.ne(1).to(torch_device)
        sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
        model = Qwen3MoeForSequenceClassification(config)
        model.to(torch_device)
        model.eval()
        result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
        self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))

    def test_Qwen3Moe_sequence_classification_model_for_single_label(self):
        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.num_labels = 3
        config.problem_type = "single_label_classification"
        input_ids = input_dict["input_ids"]
        attention_mask = input_ids.ne(1).to(torch_device)
        sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
        model = Qwen3MoeForSequenceClassification(config)
        model.to(torch_device)
        model.eval()
        result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
        self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))

    def test_Qwen3Moe_sequence_classification_model_for_multi_label(self):
        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.num_labels = 3
        config.problem_type = "multi_label_classification"
        input_ids = input_dict["input_ids"]
        attention_mask = input_ids.ne(1).to(torch_device)
        sequence_labels = ids_tensor(
            [self.model_tester.batch_size, config.num_labels], self.model_tester.type_sequence_label_size
        ).to(torch.float)
        model = Qwen3MoeForSequenceClassification(config)
        model.to(torch_device)
        model.eval()
        result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
        self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))

    # Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_llama_token_classification_model with Llama->Qwen3Moe,llama->Qwen3Moe
    def test_Qwen3Moe_token_classification_model(self):
        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.num_labels = 3
        input_ids = input_dict["input_ids"]
        attention_mask = input_ids.ne(1).to(torch_device)
        token_labels = ids_tensor([self.model_tester.batch_size, self.model_tester.seq_length], config.num_labels)
        model = Qwen3MoeForTokenClassification(config=config)
        model.to(torch_device)
        model.eval()
        result = model(input_ids, attention_mask=attention_mask, labels=token_labels)
        self.assertEqual(
            result.logits.shape,
            (self.model_tester.batch_size, self.model_tester.seq_length, self.model_tester.num_labels),
        )

    @require_flash_attn
    @require_torch_gpu
    @pytest.mark.flash_attn_test
    @slow
    def test_flash_attn_2_inference_equivalence_right_padding(self):
        self.skipTest(reason="Qwen3Moe flash attention does not support right padding")

    # Ignore copy
    def test_load_balancing_loss(self):
        r"""
        Let's make sure we can actually compute the loss and do a backward on it.
        """
        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.num_labels = 3
        config.num_experts = 8
        config.expert_interval = 2
        config.output_router_logits = True
        input_ids = input_dict["input_ids"]
        attention_mask = input_ids.ne(1).to(torch_device)
        model = Qwen3MoeForCausalLM(config)
        model.to(torch_device)
        model.eval()
        result = model(input_ids, attention_mask=attention_mask)
        self.assertEqual(result.router_logits[0].shape, (91, config.num_experts))
        torch.testing.assert_close(result.aux_loss.cpu(), torch.tensor(2, dtype=torch.float32), rtol=1e-2, atol=1e-2)

        # First, we make sure that adding padding tokens doesn't change the loss
        # loss(input_ids, attention_mask=None) == loss(input_ids + padding, attention_mask=attention_mask_with_padding)
        pad_length = 1000
        # Add padding tokens (assume that pad_token_id=1) to input_ids
        padding_block = torch.ones(input_ids.shape[0], pad_length, dtype=torch.int32).to(torch_device)
        padded_input_ids = torch.cat((padding_block, input_ids), dim=1)  # this is to simulate padding to the left
        padded_attention_mask = padded_input_ids.ne(1).to(torch_device)

        padded_result = model(padded_input_ids, attention_mask=padded_attention_mask)
        torch.testing.assert_close(result.aux_loss.cpu(), padded_result.aux_loss.cpu(), rtol=1e-4, atol=1e-4)

        # We make sure that the loss of includding padding tokens != the loss without padding tokens
        # if attention_mask=None --> we don't exclude padding tokens
        include_padding_result = model(padded_input_ids, attention_mask=None)

        # This is to mimic torch.testing.assert_not_close
        self.assertNotAlmostEqual(include_padding_result.aux_loss.item(), result.aux_loss.item())


@require_torch
class Qwen3MoeIntegrationTest(unittest.TestCase):
    @slow
    def test_model_15b_a2b_logits(self):
        input_ids = [1, 306, 4658, 278, 6593, 310, 2834, 338]
        model = Qwen3MoeForCausalLM.from_pretrained("Qwen/Qwen3-15B-A2B-Base", device_map="auto")
        input_ids = torch.tensor([input_ids]).to(model.model.embed_tokens.weight.device)
        with torch.no_grad():
            out = model(input_ids).logits.float().cpu()
        # Expected mean on dim = -1
        EXPECTED_MEAN = torch.tensor([[-1.1184, 1.1356, 9.2112, 8.0254, 5.1663, 7.9287, 8.9245, 10.0671]])
        torch.testing.assert_close(out.mean(-1), EXPECTED_MEAN, rtol=1e-2, atol=1e-2)
        # slicing logits[0, 0, 0:30]
        EXPECTED_SLICE = torch.tensor([7.5938, 2.6094, 4.0312, 4.0938, 2.5156, 2.7812, 2.9688, 1.5547, 1.3984, 2.2344, 3.0156, 3.1562, 1.1953, 3.2500, 1.0938, 8.4375, 9.5625, 9.0625, 7.5625, 7.5625, 7.9062, 7.2188, 7.0312, 6.9375, 8.0625, 1.7266, 0.9141, 3.7969, 5.3438, 3.9844])  # fmt: skip
        torch.testing.assert_close(out[0, 0, :30], EXPECTED_SLICE, rtol=1e-4, atol=1e-4)

        del model
        backend_empty_cache(torch_device)
        gc.collect()

    @slow
    def test_model_15b_a2b_generation(self):
        EXPECTED_TEXT_COMPLETION = (
            """To be or not to be, that is the question. Whether 'tis nobler in the mind to suffer the sl"""
        )
        prompt = "To be or not to"
        tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-15B-A2B-Base", use_fast=False)
        model = Qwen3MoeForCausalLM.from_pretrained("Qwen/Qwen3-15B-A2B-Base", device_map="auto")
        input_ids = tokenizer.encode(prompt, return_tensors="pt").to(model.model.embed_tokens.weight.device)

        # greedy generation outputs
        generated_ids = model.generate(input_ids, max_new_tokens=20, temperature=0)
        text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
        self.assertEqual(EXPECTED_TEXT_COMPLETION, text)

        del model
        backend_empty_cache(torch_device)
        gc.collect()

    @require_bitsandbytes
    @slow
    @require_flash_attn
    @pytest.mark.flash_attn_test
    def test_model_15b_a2b_long_prompt(self):
        EXPECTED_OUTPUT_TOKEN_IDS = [306, 338]
        # An input with 4097 tokens that is above the size of the sliding window
        input_ids = [1] + [306, 338] * 2048
        model = Qwen3MoeForCausalLM.from_pretrained(
            "Qwen/Qwen3-15B-A2B-Base",
            device_map="auto",
            load_in_4bit=True,
            attn_implementation="flash_attention_2",
        )
        input_ids = torch.tensor([input_ids]).to(model.model.embed_tokens.weight.device)
        generated_ids = model.generate(input_ids, max_new_tokens=4, temperature=0)
        self.assertEqual(EXPECTED_OUTPUT_TOKEN_IDS, generated_ids[0][-2:].tolist())

        # Assisted generation
        assistant_model = model
        assistant_model.generation_config.num_assistant_tokens = 2
        assistant_model.generation_config.num_assistant_tokens_schedule = "constant"
        generated_ids = model.generate(input_ids, max_new_tokens=4, temperature=0)
        self.assertEqual(EXPECTED_OUTPUT_TOKEN_IDS, generated_ids[0][-2:].tolist())

        del assistant_model
        del model
        backend_empty_cache(torch_device)
        gc.collect()

    @slow
    @require_torch_sdpa
    def test_model_15b_a2b_long_prompt_sdpa(self):
        EXPECTED_OUTPUT_TOKEN_IDS = [306, 338]
        # An input with 4097 tokens that is above the size of the sliding window
        input_ids = [1] + [306, 338] * 2048
        model = Qwen3MoeForCausalLM.from_pretrained(
            "Qwen/Qwen3-15B-A2B-Base",
            device_map="auto",
            attn_implementation="sdpa",
        )
        input_ids = torch.tensor([input_ids]).to(model.model.embed_tokens.weight.device)
        generated_ids = model.generate(input_ids, max_new_tokens=4, temperature=0)
        self.assertEqual(EXPECTED_OUTPUT_TOKEN_IDS, generated_ids[0][-2:].tolist())

        # Assisted generation
        assistant_model = model
        assistant_model.generation_config.num_assistant_tokens = 2
        assistant_model.generation_config.num_assistant_tokens_schedule = "constant"
        generated_ids = assistant_model.generate(input_ids, max_new_tokens=4, temperature=0)
        self.assertEqual(EXPECTED_OUTPUT_TOKEN_IDS, generated_ids[0][-2:].tolist())

        del assistant_model

        backend_empty_cache(torch_device)
        gc.collect()

        EXPECTED_TEXT_COMPLETION = (
            """To be or not to be, that is the question. Whether 'tis nobler in the mind to suffer the sl"""
        )
        prompt = "To be or not to"
        tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-15B-A2B-Base", use_fast=False)

        input_ids = tokenizer.encode(prompt, return_tensors="pt").to(model.model.embed_tokens.weight.device)

        # greedy generation outputs
        generated_ids = model.generate(input_ids, max_new_tokens=20, temperature=0)
        text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
        self.assertEqual(EXPECTED_TEXT_COMPLETION, text)

    @slow
    def test_speculative_generation(self):
        EXPECTED_TEXT_COMPLETION = (
            "To be or not to be, that is the question: whether 'tis nobler in the mind to suffer the sl"
        )
        prompt = "To be or not to"
        tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-15B-A2B-Base", use_fast=False)
        model = Qwen3MoeForCausalLM.from_pretrained(
            "Qwen/Qwen3-15B-A2B-Base", device_map="auto", torch_dtype=torch.float16
        )
        assistant_model = Qwen3MoeForCausalLM.from_pretrained(
            "Qwen/Qwen3-15B-A2B-Base", device_map="auto", torch_dtype=torch.float16
        )
        input_ids = tokenizer.encode(prompt, return_tensors="pt").to(model.model.embed_tokens.weight.device)

        # greedy generation outputs
        set_seed(0)
        generated_ids = model.generate(
            input_ids, max_new_tokens=20, do_sample=True, temperature=0.3, assistant_model=assistant_model
        )
        text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
        self.assertEqual(EXPECTED_TEXT_COMPLETION, text)

        del model
        backend_empty_cache(torch_device)
        gc.collect()