A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 7, 14:00 UTC
[MTP] Validate that MTP weights are actually loaded (#35548)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Feb 27, 2026 at 23:27 UTC
2562e0271e51950da23c57b237649250bf41d9e0
1 file changed
+20
vllm/model_executor/models/deepseek_mtp.py
+20
@@ -415,6 +415,26 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
415
weight_loader(param, loaded_weight)
416
if not is_fusion_moe_shared_experts_layer:
417
loaded_params.add(name)
418
+
419
+ # Validate that weights were loaded for each expected MTP layer.
420
+ loaded_layers: set[int] = set()
421
+ for param_name in loaded_params:
422
+ spec_layer = get_spec_layer_idx_from_weight_name(self.config, param_name)
423
+ if spec_layer is not None:
424
+ loaded_layers.add(spec_layer)
425
+ for layer_idx in range(
426
+ self.model.mtp_start_layer_idx,
427
+ self.model.mtp_start_layer_idx + self.model.num_mtp_layers,
428
+ ):
429
+ if layer_idx not in loaded_layers:
430
+ raise ValueError(
431
+ f"MTP speculative decoding layer {layer_idx} weights "
432
+ f"missing from checkpoint. The checkpoint may have "
433
+ f"been quantized without including the MTP layers. "
434
+ f"Use a checkpoint that includes MTP layer weights, "
435
+ f"or disable speculative decoding."
436
+ )
437
+
438
return loaded_params
439
440
def _rewrite_spec_layer_name(self, spec_layer: int, name: str) -> str: