[Bugfix][Model][Spec Decode] Defer disposable GLM MTP head (#55442)

Signed-off-by: Luca Motz <luca.motz@icloud.com>
This commit is contained in:
lucamotz
2026-09-22 20:08:27 +00:00
committed by GitHub
parent 71a2392845
commit 276db94fb2
3 changed files with 40 additions and 8 deletions
+26
View File
@@ -5,6 +5,7 @@ from unittest import mock
import pytest
import torch
import torch.nn as nn
from tests.v1.attention.utils import (
BatchSpec,
@@ -31,6 +32,31 @@ mimo_7b_dir = "XiaomiMiMo/MiMo-7B-Base"
DEVICE_TYPE = current_platform.device_type
def test_glm_mtp_defers_lm_head(default_vllm_config):
from vllm.model_executor.models import deepseek_mtp
from vllm.models.glm5next.common import mtp
config = mock.MagicMock(
hidden_size=16,
rms_norm_eps=1e-5,
index_topk=8,
index_kpool=4,
)
vllm_config = mock.MagicMock()
vllm_config.speculative_config.draft_model_config.hf_config = config
vllm_config.scheduler_config.max_num_batched_tokens = 4
with (
mock.patch.object(mtp, "Glm5NextDecoderLayer", return_value=nn.Identity()),
mock.patch.object(deepseek_mtp, "ParallelLMHead") as parallel_lm_head,
mock.patch.object(mtp.current_platform, "device_type", "cpu"),
):
layer = mtp.Glm5NextMultiTokenPredictorLayer(vllm_config, "model.layers.1")
parallel_lm_head.assert_not_called()
assert layer.shared_head.head is None
def _create_mtp_proposer(num_speculative_tokens: int) -> EagleProposer:
"""Create an MTP proposer with unified model configuration."""
model_config = ModelConfig(