[Bugfix][RL] Load attention sinks through weight_loader for layerwise reload (#57797)

Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
Co-authored-by: aoshen02 <aoshen@inferact.ai>
This commit is contained in:
Ronald
2026-10-14 10:17:58 +00:00
committed by GitHub
co-authored by aoshen02
parent f9b65c30c7
commit 19aba3e4d5
30 changed files with 403 additions and 476 deletions
@@ -42,10 +42,6 @@ def _disable_distributed_loader_paths(monkeypatch: pytest.MonkeyPatch):
"fused_moe_make_expert_params_mapping",
lambda *args, **kwargs: [],
)
monkeypatch.setattr(
dspark_module, "get_tensor_model_parallel_world_size", lambda: 1
)
monkeypatch.setattr(dspark_module, "get_tensor_model_parallel_rank", lambda: 0)
@pytest.mark.cpu_test
-4
View File
@@ -258,8 +258,6 @@ def test_v41_dspark_loads_linear_scales(
draft._remap_dspark_name = lambda name: (
dspark.DSparkDeepseekV4ForCausalLM._remap_dspark_name(draft, name)
)
monkeypatch.setattr(dspark, "get_tensor_model_parallel_world_size", lambda: 4)
monkeypatch.setattr(dspark, "get_tensor_model_parallel_rank", lambda: 0)
monkeypatch.setattr(
dspark, "fused_moe_make_expert_params_mapping", lambda *a, **kw: []
)
@@ -310,8 +308,6 @@ def test_dsv4_dspark_fused_moe_expert_mapping_includes_redundant_experts(
draft._remap_dspark_name = lambda name: (
dspark.DSparkDeepseekV4ForCausalLM._remap_dspark_name(draft, name)
)
monkeypatch.setattr(dspark, "get_tensor_model_parallel_world_size", lambda: 1)
monkeypatch.setattr(dspark, "get_tensor_model_parallel_rank", lambda: 0)
monkeypatch.setattr(
dspark,
"fused_moe_make_expert_params_mapping",