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 17, 01:45 UTC
Sync huggingface modifications of qwen Moe model (#4774)
eigenLiu committed
May 18, 2024 at 00:43 UTC
48d5985a088c6e13e9ad9b0c7a0ce846e30b529f
1 file changed
+9
-2
vllm/model_executor/models/qwen2_moe.py
+9
-2
@@ -283,8 +283,9 @@ class Qwen2MoeDecoderLayer(nn.Module):
283
cache_config=cache_config,
284
quant_config=quant_config,
285
)
286
- if (config.num_experts is not None
287
- and (layer_idx + 1) % config.decoder_sparse_step == 0):
286
+ if (layer_idx not in config.mlp_only_layers) and (
287
+ config.num_experts > 0 and
288
+ (layer_idx + 1) % config.decoder_sparse_step == 0):
289
self.mlp = Qwen2MoeSparseMoeBlock(config=config,
290
quant_config=quant_config)
291
else:
@@ -439,6 +440,9 @@ class Qwen2MoeForCausalLM(nn.Module):
440
if (("mlp.experts." in name or "mlp.shared_expert." in name)
441
and name not in params_dict):
442
continue
443
+ if name not in params_dict:
444
+ continue
445
+
446
param = params_dict[name]
447
weight_loader = param.weight_loader
448
weight_loader(param, loaded_weight, shard_id)
@@ -451,6 +455,9 @@ class Qwen2MoeForCausalLM(nn.Module):
455
if (("mlp.experts." in name or "mlp.shared_expert." in name)
456
and name not in params_dict):
457
continue
458
+ if name not in params_dict:
459
+ continue
460
+
461
param = params_dict[name]
462
weight_loader = getattr(param, "weight_loader",
463
default_weight_loader)