Merge branch 'main' into patches

Signed-off-by: MOVZX <movzx@yahoo.com>
This commit is contained in:
2026-10-16 17:35:42 +07:00
347 changed files with 29531 additions and 3472 deletions
+30 -13
View File
@@ -2132,8 +2132,8 @@ def test_draft_parallel_config_preserves_ep_without_model(
[
pytest.param("eagle3", False, 0, id="eagle3"),
pytest.param("eagle3", True, 7, id="p-eagle"),
pytest.param("dflash", True, 8, id="dflash"),
pytest.param("dspark", True, 7, id="dspark"),
pytest.param("dflash", True, 0, id="dflash"),
pytest.param("dspark", True, 0, id="dspark"),
pytest.param("mtp", False, 0, id="mtp"),
pytest.param("ngram", False, 0, id="ngram"),
pytest.param("draft_model", False, 1, id="draft-model"),
@@ -3769,34 +3769,48 @@ def test_dual_key_gumbel_requires_probabilistic_drafting():
config._check_supports_watermarking()
@pytest.mark.parametrize("method", ["eagle", "eagle3", "mtp"])
def test_dual_key_gumbel_supports_probabilistic_speculative_decoding(method):
@pytest.mark.parametrize(
("method", "parallel_drafting"),
[
("eagle", False),
("eagle3", False),
("mtp", False),
("dspark", True),
("dflash", True),
],
)
def test_dual_key_gumbel_supports_probabilistic_speculative_decoding(
method, parallel_drafting
):
config = _watermarked_vllm_config()
config.watermark_config = WatermarkConfig(algorithm="dual_key_gumbel", key=42)
config.speculative_config = SimpleNamespace(
method=method,
draft_sample_method="probabilistic",
rejection_sample_method="standard",
parallel_drafting=False,
parallel_drafting=parallel_drafting,
)
config._check_supports_watermarking()
def test_dual_key_gumbel_supports_dspark():
@pytest.mark.parametrize("architecture", ["DFlash2DraftModel", "LiLiCorrDraftModel"])
def test_dual_key_gumbel_rejects_dflash_candidate_drafts(architecture):
config = _watermarked_vllm_config()
config.watermark_config = WatermarkConfig(algorithm="dual_key_gumbel", key=42)
config.speculative_config = SimpleNamespace(
method="dspark",
method="dflash",
draft_sample_method="probabilistic",
rejection_sample_method="standard",
parallel_drafting=True,
draft_model_config=SimpleNamespace(architectures=[architecture]),
)
config._check_supports_watermarking()
with pytest.raises(ValueError, match="candidate-head drafters"):
config._check_supports_watermarking()
def test_dual_key_gumbel_rejects_non_autoregressive_speculation():
def test_dual_key_gumbel_rejects_unsupported_speculative_method():
config = _watermarked_vllm_config()
config.watermark_config = WatermarkConfig(algorithm="dual_key_gumbel", key=42)
config.speculative_config = SimpleNamespace(
@@ -3806,7 +3820,7 @@ def test_dual_key_gumbel_rejects_non_autoregressive_speculation():
parallel_drafting=False,
)
with pytest.raises(ValueError, match="autoregressive model-based"):
with pytest.raises(ValueError, match="Watermarking supports only"):
config._check_supports_watermarking()
@@ -4167,7 +4181,7 @@ def test_revision_resolved_for_model(mock_resolve):
@pytest.mark.parametrize(
"enabled,supported,is_cuda,tp_size,pp_size,error",
"enabled,supported,cuda_alike,tp_size,pp_size,error",
[
(True, True, True, 2, 1, None),
(True, False, True, 2, 1, "not supported for this model"),
@@ -4178,7 +4192,7 @@ def test_revision_resolved_for_model(mock_resolve):
(False, False, False, 2, 1, None),
],
)
def test_hc_sp_model_support(enabled, supported, is_cuda, tp_size, pp_size, error):
def test_hc_sp_model_support(enabled, supported, cuda_alike, tp_size, pp_size, error):
"""Validate active HC SP configurations and ignore the flag with TP=1."""
from vllm.model_executor.models.registry import ModelRegistry
@@ -4196,7 +4210,10 @@ def test_hc_sp_model_support(enabled, supported, is_cuda, tp_size, pp_size, erro
distributed_executor_backend="mp",
)
with (
patch("vllm.config.model.current_platform.is_cuda", return_value=is_cuda),
patch(
"vllm.config.model.current_platform.is_cuda_alike",
return_value=cuda_alike,
),
patch.object(ModelRegistry, "is_hc_sp_supported_model", return_value=supported),
patch.object(ModelRegistry, "is_pp_supported_model", return_value=True),
):