[Bugfix][Spec Decode] Separate DSpark width from MTP stage validation (#54631)

Signed-off-by: Luca Motz <luca.motz@icloud.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
lucamotz
2026-09-24 06:08:22 +00:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 44bf011f44
commit ee2f3cdf12
2 changed files with 96 additions and 16 deletions
@@ -47,6 +47,91 @@ def test_none_overrides_fall_back_to_arch_mapping():
assert composed is SpeculativeConfig.hf_config_override
def _make_speculative_config(
hf_kwargs: dict, architecture: str, **speculative_kwargs
) -> SpeculativeConfig:
hf_config = SpeculativeConfig.hf_config_override(_make_hf_config(**hf_kwargs))
draft = MagicMock(
hf_config=hf_config, architectures=hf_config.architectures, max_model_len=128
)
draft.registry.inspect_model_cls.return_value = (None, architecture)
target = MagicMock(max_model_len=128, quantization=None, hf_overrides={})
target.hf_config.model_type = hf_kwargs["model_type"]
with patch("vllm.config.speculative.ModelConfig", return_value=draft):
return SpeculativeConfig(
model="draft",
**speculative_kwargs,
target_model_config=target,
target_parallel_config=ParallelConfig(),
)
@pytest.mark.cpu_test
@pytest.mark.parametrize(
("hf_kwargs", "n_predict", "architecture"),
[
(
dict(
model_type="deepseek_v4",
num_nextn_predict_layers=3,
dspark_block_size=5,
),
3,
"DSparkDraftModel",
),
(
dict(
model_type="deepseek_v41",
num_nextn_predict_layers=3,
dspark_block_size=5,
),
3,
"DSparkV41DraftModel",
),
(
dict(
model_type="gemma4_text",
architectures=["Gemma4DSparkModel"],
block_size=5,
),
None,
"Gemma4DSparkModel",
),
],
)
@pytest.mark.parametrize("speculative_kwargs", [{"num_speculative_tokens": 5}, {}])
def test_dspark_width_is_independent_of_mtp_stages(
hf_kwargs, n_predict, architecture, speculative_kwargs
):
config = _make_speculative_config(
hf_kwargs, architecture, method="dspark", **speculative_kwargs
)
assert config.num_speculative_tokens == 5
assert getattr(config.draft_model_config.hf_config, "n_predict", None) == n_predict
assert config.draft_model_config.hf_config.architectures == [architecture]
@pytest.mark.cpu_test
def test_mtp_stages_are_independent_of_dspark_width():
hf_kwargs = dict(
model_type="deepseek_v4", num_nextn_predict_layers=3, dspark_block_size=5
)
config = _make_speculative_config(
hf_kwargs, "DeepSeekV4MTPModel", method="mtp", num_speculative_tokens=6
)
assert config.num_speculative_tokens == 6
assert config.draft_model_config.hf_config.n_predict == 3
assert config.draft_model_config.hf_config.architectures == ["DeepSeekV4MTPModel"]
with pytest.raises(ValueError, match="must be divisible by n_predict=3"):
_make_speculative_config(
hf_kwargs, "DeepSeekV4MTPModel", method="mtp", num_speculative_tokens=5
)
@pytest.mark.cpu_test
def test_callable_overrides_reach_the_draft_config():
"""A callable override (config-to-config transform) composes with the