[Bugfix] Fix DSV32 weight loading (#38870)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
This commit is contained in:
Yongye Zhu
2026-04-03 19:57:52 -07:00
committed by GitHub
parent 06fd9ffcc4
commit 8617f8676b
2 changed files with 68 additions and 27 deletions
+13 -3
View File
@@ -184,11 +184,16 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
self.config = vllm_config.model_config.hf_config
self.quant_config = vllm_config.quant_config
self.model = DeepSeekMultiTokenPredictor(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
# Set MoE hyperparameters
self.set_moe_parameters()
self.is_fp4_ckpt = (
self.quant_config is not None
and self.quant_config.get_name() == "modelopt_fp4"
)
def set_moe_parameters(self):
self.expert_weights = []
@@ -241,11 +246,16 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
("gate_up_proj", "up_proj", 1),
("fused_qkv_a_proj", "q_a_proj", 0),
("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1),
# Fused indexer wk + weights_proj
("wk_weights_proj", "wk", 0),
("wk_weights_proj", "weights_proj", 1),
]
if self.is_fp4_ckpt:
# Fused indexer wk + weights_proj (shard 0 = wk, shard 1 = weights_proj)
indexer_fused_mapping = [
("wk_weights_proj", "wk", 0),
("wk_weights_proj", "weights_proj", 1),
]
stacked_params_mapping.extend(indexer_fused_mapping)
expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
self,
ckpt_gate_proj_name="gate_proj",