[Model] Add Qwen3-Omni DSpark support (#52560)
Signed-off-by: Zhou248 <630563665@qq.com>
This commit is contained in:
@@ -32,6 +32,7 @@ from vllm.config.compilation import CompilationMode, CUDAGraphMode
|
||||
from vllm.config.kernel import IrOpPriorityConfig
|
||||
from vllm.config.load import LoadConfig
|
||||
from vllm.config.mamba import MambaBackendEnum
|
||||
from vllm.config.speculative import _validate_qwen3_omni_dspark
|
||||
from vllm.config.utils import get_field
|
||||
from vllm.config.vllm import OPTIMIZATION_LEVEL_TO_CONFIG, OptimizationLevel
|
||||
from vllm.platforms import current_platform
|
||||
@@ -2100,6 +2101,126 @@ def test_mtp_draft_uses_model_weights_not_local_cache(mock_model_config_cls):
|
||||
assert actual_model == s3_url
|
||||
|
||||
|
||||
def _make_qwen3_omni_dspark_configs():
|
||||
text_config = SimpleNamespace(
|
||||
hidden_size=2048,
|
||||
num_hidden_layers=48,
|
||||
num_attention_heads=32,
|
||||
num_key_value_heads=4,
|
||||
head_dim=128,
|
||||
vocab_size=152064,
|
||||
)
|
||||
target_model_config = SimpleNamespace(
|
||||
hf_config=SimpleNamespace(
|
||||
model_type="qwen3_omni_moe",
|
||||
architectures=["Qwen3OmniMoeForConditionalGeneration"],
|
||||
thinker_config=SimpleNamespace(text_config=text_config),
|
||||
),
|
||||
hf_text_config=text_config,
|
||||
architectures=["Qwen3OmniMoeForConditionalGeneration"],
|
||||
get_hidden_size=lambda: 2048,
|
||||
get_total_num_hidden_layers=lambda: 48,
|
||||
get_vocab_size=lambda: 152064,
|
||||
)
|
||||
draft_hf_config = SimpleNamespace(
|
||||
model_type="qwen3",
|
||||
architectures=["Qwen3OmniDSparkModel"],
|
||||
hidden_size=2048,
|
||||
num_attention_heads=32,
|
||||
num_key_value_heads=4,
|
||||
head_dim=128,
|
||||
block_size=7,
|
||||
target_hidden_size=2048,
|
||||
target_layer_ids=[1, 9, 17, 25, 33],
|
||||
use_aux_hidden_state=True,
|
||||
markov_rank=64,
|
||||
markov_head_type="vanilla",
|
||||
sample_from_anchor=True,
|
||||
dspark_bonus_anchor=False,
|
||||
vocab_size=152064,
|
||||
draft_vocab_size=32000,
|
||||
mask_token_id=151669,
|
||||
rope_parameters={"rope_type": "default", "rope_theta": 1000000.0},
|
||||
)
|
||||
draft_model_config = SimpleNamespace(
|
||||
hf_config=draft_hf_config,
|
||||
architectures=["Qwen3OmniDSparkModel"],
|
||||
)
|
||||
return target_model_config, draft_model_config
|
||||
|
||||
|
||||
@pytest.mark.skip_global_cleanup
|
||||
def test_qwen3_omni_dspark_checkpoint_contract_is_accepted():
|
||||
target_config, draft_config = _make_qwen3_omni_dspark_configs()
|
||||
_validate_qwen3_omni_dspark(target_config, draft_config, 7)
|
||||
|
||||
|
||||
@pytest.mark.skip_global_cleanup
|
||||
def test_qwen3_omni_dspark_rejects_generic_qwen3_architecture():
|
||||
target_config, draft_config = _make_qwen3_omni_dspark_configs()
|
||||
draft_config.hf_config.architectures = ["Qwen3DSparkModel"]
|
||||
draft_config.architectures = ["Qwen3DSparkModel"]
|
||||
|
||||
with pytest.raises(ValueError, match="must be converted first"):
|
||||
_validate_qwen3_omni_dspark(target_config, draft_config, 7)
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
("field", "value", "error"),
|
||||
[
|
||||
("block_size", 5, "trained block_size"),
|
||||
("target_hidden_size", 4096, "target_hidden_size"),
|
||||
("hidden_size", 4096, "draft hidden_size"),
|
||||
("num_attention_heads", 16, "num_attention_heads"),
|
||||
("num_key_value_heads", 8, "num_key_value_heads"),
|
||||
("head_dim", 64, "head_dim"),
|
||||
("target_layer_ids", [7, 48], "zero-based text-layer"),
|
||||
("target_layer_ids", [23, 7], "strictly increasing"),
|
||||
("use_aux_hidden_state", False, "use_aux_hidden_state=true"),
|
||||
("markov_rank", 0, "markov_rank"),
|
||||
("markov_head_type", "gated", "markov_head_type='vanilla'"),
|
||||
("sample_from_anchor", False, "sample_from_anchor=true"),
|
||||
("dspark_bonus_anchor", True, "dspark_bonus_anchor=false"),
|
||||
("vocab_size", 0, "input vocab_size must be a positive integer"),
|
||||
],
|
||||
)
|
||||
@pytest.mark.skip_global_cleanup
|
||||
def test_qwen3_omni_dspark_rejects_incompatible_checkpoint_fields(field, value, error):
|
||||
target_config, draft_config = _make_qwen3_omni_dspark_configs()
|
||||
setattr(draft_config.hf_config, field, value)
|
||||
|
||||
with pytest.raises(ValueError, match=error):
|
||||
_validate_qwen3_omni_dspark(target_config, draft_config, 7)
|
||||
|
||||
|
||||
@pytest.mark.skip_global_cleanup
|
||||
def test_qwen3_omni_dspark_rejects_mrope_draft_positions():
|
||||
target_config, draft_config = _make_qwen3_omni_dspark_configs()
|
||||
draft_config.hf_config.rope_parameters["mrope_section"] = [24, 20, 20]
|
||||
|
||||
with pytest.raises(ValueError, match="logical 1-D RoPE"):
|
||||
_validate_qwen3_omni_dspark(target_config, draft_config, 7)
|
||||
|
||||
|
||||
@pytest.mark.skip_global_cleanup
|
||||
def test_qwen3_omni_dspark_allows_draft_only_noise_token_row():
|
||||
target_config, draft_config = _make_qwen3_omni_dspark_configs()
|
||||
draft_config.hf_config.vocab_size = 152065
|
||||
draft_config.hf_config.draft_vocab_size = 152064
|
||||
draft_config.hf_config.mask_token_id = 152064
|
||||
|
||||
_validate_qwen3_omni_dspark(target_config, draft_config, 7)
|
||||
|
||||
|
||||
@pytest.mark.skip_global_cleanup
|
||||
def test_qwen3_omni_dspark_allows_smaller_input_vocabulary():
|
||||
target_config, draft_config = _make_qwen3_omni_dspark_configs()
|
||||
draft_config.hf_config.vocab_size = 151936
|
||||
draft_config.hf_config.mask_token_id = 151669
|
||||
|
||||
_validate_qwen3_omni_dspark(target_config, draft_config, 7)
|
||||
|
||||
|
||||
def test_ir_op_priority_default():
|
||||
"""Test that IR op priority defaults are set correctly."""
|
||||
from vllm.config.kernel import IrOpPriorityConfig
|
||||
|
||||
Reference in New Issue
Block a user