move apply_torchao_config_ to model_runner (#2342)

python/sglang/srt/models/grok.py

@@ -35,12 +35,10 @@ from sglang.srt.layers.linear import (
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.layers.torchao_utils import apply_torchao_config_
 from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.loader import DefaultModelLoader
 from sglang.srt.model_loader.weight_utils import default_weight_loader
@@ -290,7 +288,6 @@ class Grok1ForCausalLM(nn.Module):
         super().__init__()
         self.config = config
         self.quant_config = quant_config
-        self.torchao_config = global_server_args_dict["torchao_config"]
         self.model = Grok1Model(config, quant_config=quant_config)
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
@@ -374,8 +371,6 @@ class Grok1ForCausalLM(nn.Module):
                     )
                     weight_loader(param, loaded_weight)
 
-        apply_torchao_config_(self, params_dict, set(["proj.weight"]))
-
 
 class Grok1ModelForCausalLM(Grok1ForCausalLM):
     """An alias for backward-compatbility."""

python/sglang/srt/models/llama.py

@@ -36,12 +36,10 @@ from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorO
 from sglang.srt.layers.pooler import Pooler, PoolingType
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.layers.torchao_utils import apply_torchao_config_
 from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import make_layers
@@ -304,7 +302,6 @@ class LlamaForCausalLM(nn.Module):
         super().__init__()
         self.config = config
         self.quant_config = quant_config
-        self.torchao_config = global_server_args_dict["torchao_config"]
         self.model = LlamaModel(config, quant_config=quant_config)
         # Llama 3.2 1B Insturct set tie_word_embeddings to True
         # Llama 3.1 8B Insturct set tie_word_embeddings to False
@@ -424,8 +421,6 @@ class LlamaForCausalLM(nn.Module):
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)
 
-        apply_torchao_config_(self, params_dict, set(["proj.weight"]))
-
     def get_weights_by_name(
         self, name: str, truncate_size: int = 100, tp_size: int = 1
     ) -> Optional[torch.Tensor]:

python/sglang/srt/models/mixtral.py

@@ -34,12 +34,10 @@ from sglang.srt.layers.linear import (
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.layers.torchao_utils import apply_torchao_config_
 from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 
@@ -295,7 +293,6 @@ class MixtralForCausalLM(nn.Module):
         super().__init__()
         self.config = config
         self.quant_config = quant_config
-        self.torchao_config = global_server_args_dict["torchao_config"]
         self.model = MixtralModel(config, quant_config=quant_config, prefix="model")
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
@@ -387,7 +384,5 @@ class MixtralForCausalLM(nn.Module):
                     )
                     weight_loader(param, loaded_weight)
 
-        apply_torchao_config_(self, params_dict, set(["proj.weight"]))
-
 
 EntryClass = MixtralForCausalLM

python/sglang/srt/models/phi3_small.py

@@ -17,13 +17,11 @@ from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorO
 from sglang.srt.layers.pooler import Pooler, PoolingType
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.layers.torchao_utils import apply_torchao_config_
 from sglang.srt.layers.vocab_parallel_embedding import (
     DEFAULT_VOCAB_PADDING_SIZE,
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import make_layers
@@ -348,7 +346,6 @@ class Phi3SmallForCausalLM(nn.Module):
             quant_config=quant_config,
             prefix="model",
         )
-        self.torchao_config = global_server_args_dict["torchao_config"]
         self.vocab_size = config.vocab_size
         self.mup_width_multiplier = config.mup_width_multiplier
         self.lm_head = ParallelLMHead(
@@ -441,7 +438,5 @@ class Phi3SmallForCausalLM(nn.Module):
             weight_loader = getattr(param, "weight_loader", default_weight_loader)
             weight_loader(param, loaded_weight)
 
-        apply_torchao_config_(self, params_dict, set(["proj.weight"]))
-
 
 EntryClass = Phi3SmallForCausalLM

python/sglang/srt/models/qwen2_moe.py

@@ -40,12 +40,10 @@ from sglang.srt.layers.linear import (
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.layers.torchao_utils import apply_torchao_config_
 from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 
@@ -352,7 +350,6 @@ class Qwen2MoeForCausalLM(nn.Module):
         super().__init__()
         self.config = config
         self.quant_config = quant_config
-        self.torchao_config = global_server_args_dict["torchao_config"]
         self.model = Qwen2MoeModel(config, quant_config)
         self.lm_head = ParallelLMHead(
             config.vocab_size, config.hidden_size, quant_config=quant_config
@@ -445,7 +442,5 @@ class Qwen2MoeForCausalLM(nn.Module):
                     )
                     weight_loader(param, loaded_weight)
 
-        apply_torchao_config_(self, params_dict, set(["proj.weight"]))
-
 
 EntryClass = Qwen2MoeForCausalLM

python/sglang/srt/models/torch_native_llama.py

@@ -58,12 +58,10 @@ from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.layers.torchao_utils import apply_torchao_config_
 from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 
@@ -392,7 +390,6 @@ class TorchNativeLlamaForCausalLM(nn.Module):
         super().__init__()
         self.config = config
         self.quant_config = quant_config
-        self.torchao_config = global_server_args_dict["torchao_config"]
         self.supports_torch_tp = True
         self.model = LlamaModel(config, quant_config=quant_config)
         if self.config.tie_word_embeddings:
@@ -503,8 +500,6 @@ class TorchNativeLlamaForCausalLM(nn.Module):
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)
 
-        apply_torchao_config_(self, params_dict, set(["proj.weight"]))
-
 
 class TorchNativePhi3ForCausalLM(TorchNativeLlamaForCausalLM):
     pass