[Bugfix] Fix DeepSeek V4 mHC broadcast buffer for dummy load (#51368)

Signed-off-by: Hollow Man <hollowman@opensuse.org>
This commit is contained in:
ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟
2026-08-18 17:02:18 -07:00
committed by GitHub
parent ef47a897e2
commit a9f4afb66f
4 changed files with 46 additions and 7 deletions
+34
View File
@@ -9,10 +9,13 @@ import torch
from vllm.model_executor.layers.fused_moe.routed_experts_capturer import (
bind_routed_experts_capturer,
)
from vllm.models.deepseek_v4.nvidia.dspark import DSparkDeepseekV4ForCausalLM
from vllm.models.deepseek_v4.nvidia.model import (
DeepseekV4ForCausalLM,
DeepseekV4MegaMoEExperts,
make_deepseek_v4_expert_params_mapping,
)
from vllm.models.deepseek_v4.nvidia.mtp import DeepSeekV4MTP
from vllm.models.deepseek_v4.nvidia.ops.prepare_megamoe import prepare_megamoe_inputs
from vllm.platforms import current_platform
@@ -243,6 +246,37 @@ def test_deepseek_v4_mega_moe_fused_input_staging_is_bitwise_exact():
)
def test_deepseek_v4_pwal_hook_finalizes_mega_moe_and_mhc_broadcast():
"""The loader invokes the model-level PWAL hook for every load format,
so it must finalize megamoe + mhc broadcast weights to cover dummy
load, which skips load_weights()."""
calls = []
stub = SimpleNamespace(
model=SimpleNamespace(
finalize_mega_moe_weights=lambda: calls.append("mega_moe"),
finalize_mhc_broadcast_weights=lambda: calls.append("mhc"),
)
)
DeepseekV4ForCausalLM.process_weights_after_loading(stub)
assert calls == ["mega_moe", "mhc"]
def test_deepseek_v4_drafter_pwal_hooks_finalize_mega_moe():
"""MTP/DSpark drafters load as their own top-level models, so each needs
its own PWAL hook now that the megamoe forward no longer finalizes
weights lazily on first use."""
calls = []
mtp = SimpleNamespace(finalize_mega_moe_weights=lambda: calls.append("mtp"))
DeepSeekV4MTP.process_weights_after_loading(mtp)
dspark = SimpleNamespace(_finalize_moe=lambda: calls.append("dspark"))
DSparkDeepseekV4ForCausalLM.process_weights_after_loading(dspark)
assert calls == ["mtp", "dspark"]
@pytest.mark.skipif(
not torch.cuda.is_available(),
reason="DeepSeek V4 MegaMoE fused input staging requires CUDA.",