test: enforce early CP HiCache storage rejection
This commit is contained in:
@@ -383,22 +383,39 @@ class TestHiCacheArgs(CustomTestCase):
|
||||
hicache_storage_backend="mooncake",
|
||||
tp_size=8,
|
||||
)
|
||||
pre_call_state = (
|
||||
args.enable_dp_attention,
|
||||
args.moe_dense_tp_size,
|
||||
args.moe_a2a_backend,
|
||||
args.ep_size,
|
||||
args.attn_cp_size,
|
||||
args.page_size,
|
||||
)
|
||||
|
||||
with self.assertRaisesRegex(
|
||||
AssertionError,
|
||||
"enable_nsa_prefill_cp_shared_kv.*hicache_storage_backend",
|
||||
with (
|
||||
self.assertRaisesRegex(
|
||||
AssertionError,
|
||||
"enable_nsa_prefill_cp_shared_kv.*hicache_storage_backend",
|
||||
),
|
||||
patch.object(
|
||||
args,
|
||||
"get_model_config",
|
||||
side_effect=AssertionError("get_model_config should not be called"),
|
||||
),
|
||||
):
|
||||
model_config = MagicMock()
|
||||
model_config.hf_config.architectures = ["DeepseekV3ForCausalLM"]
|
||||
with (
|
||||
patch.object(args, "get_model_config", return_value=model_config),
|
||||
patch(
|
||||
"sglang.srt.configs.model_config.is_deepseek_nsa",
|
||||
return_value=True,
|
||||
),
|
||||
patch("torch.cuda.get_device_capability", return_value=(9, 0)),
|
||||
):
|
||||
args._handle_model_specific_adjustments()
|
||||
args._handle_model_specific_adjustments()
|
||||
|
||||
self.assertEqual(
|
||||
pre_call_state,
|
||||
(
|
||||
args.enable_dp_attention,
|
||||
args.moe_dense_tp_size,
|
||||
args.moe_a2a_backend,
|
||||
args.ep_size,
|
||||
args.attn_cp_size,
|
||||
args.page_size,
|
||||
),
|
||||
)
|
||||
|
||||
def test_hicache_io_backend_and_mem_layout_compatibility(self):
|
||||
cases = [
|
||||
|
||||
Reference in New Issue
Block a user