Support loading of larger models with on-the-fly quantization (#3061)

python/sglang/srt/configs/load_config.py

@@ -20,6 +20,7 @@ class LoadFormat(str, enum.Enum):
     GGUF = "gguf"
     BITSANDBYTES = "bitsandbytes"
     MISTRAL = "mistral"
+    LAYERED = "layered"
 
 
 @dataclass

python/sglang/srt/layers/torchao_utils.py

@@ -5,6 +5,7 @@ Common utilities for torchao.
 import logging
 import os
 import pwd
+from typing import Callable, Optional
 
 import torch
 
@@ -27,8 +28,18 @@ def save_gemlite_cache(print_error: bool = False) -> bool:
     return True
 
 
+def proj_filter(
+    module: torch.nn.Module,
+    fqn: str,
+):
+    """Filter function for quantizing projection layers."""
+    return "proj" in fqn
+
+
 def apply_torchao_config_to_model(
-    model: torch.nn.Module, torchao_config: str, filter_fn=None
+    model: torch.nn.Module,
+    torchao_config: str,
+    filter_fn: Optional[Callable] = proj_filter,
 ):
     """Quantize a modelwith torchao quantization specified by torchao_config
 
@@ -49,11 +60,6 @@ def apply_torchao_config_to_model(
     )
     from torchao.quantization.observer import PerRow, PerTensor
 
-    if filter_fn is None:
-
-        def filter_fn(module, fqn):
-            return "proj" in fqn
-
     if torchao_config == "" or torchao_config is None:
         return model
     elif "int8wo" in torchao_config:

python/sglang/srt/model_executor/model_runner.py

@@ -185,9 +185,12 @@ class ModelRunner:
         self.load_model()
 
         # Apply torchao quantization
-        apply_torchao_config_to_model(
-            self.model, global_server_args_dict["torchao_config"]
-        )
+        torchao_applied = getattr(self.model, "torchao_applied", False)
+        # In layered loading, torchao may have been applied
+        if not torchao_applied:
+            apply_torchao_config_to_model(
+                self.model, global_server_args_dict["torchao_config"]
+            )
 
         # Apply torch TP if the model supports it
         supports_torch_tp = getattr(self.model, "supports_torch_tp", False)

python/sglang/srt/model_loader/loader.py

@@ -374,6 +374,78 @@ class DefaultModelLoader(BaseModelLoader):
         return model.eval()
 
 
+class LayeredModelLoader(DefaultModelLoader):
+    """Model loader that loads weights layer by layer so that one can quantize a
+    layer before loading another to make the peak memory envelope smaller."""
+
+    def __init__(self, load_config: LoadConfig):
+        # Back to the default load format
+        load_config.load_format = LoadFormat.AUTO
+        super().__init__(load_config)
+
+    def load_model(
+        self,
+        *,
+        model_config: ModelConfig,
+        device_config: DeviceConfig,
+    ) -> nn.Module:
+        from sglang.srt.layers.torchao_utils import apply_torchao_config_to_model
+        from sglang.srt.managers.schedule_batch import global_server_args_dict
+
+        torchao_config = global_server_args_dict.get("torchao_config")
+        target_device = torch.device(device_config.device)
+
+        with set_default_torch_dtype(model_config.dtype):
+            # Create model on meta device
+            with torch.device("meta"):
+                model = _initialize_model(
+                    model_config,
+                    self.load_config,
+                )
+
+            # Check model's layered load support
+            if not hasattr(model, "load_weights_to_module"):
+                raise ValueError(
+                    "LayeredModelLoader requires the model to have a "
+                    "`load_weights_to_module` method. "
+                    f"{model_config.model_path} does not support it."
+                )
+
+            # Get all weights from disk
+            weights = self._get_all_weights(model_config, model)
+
+            # Helper function to recursively fill the weights of a module
+            def fill_module(module, fqn: List[str], weights):
+                """
+                fqn: list of strings representing the fully qualified name of `module`.
+                """
+                # Layer by layer
+                for name, submod in module.named_children():
+                    fill_module(submod, fqn + [name], weights)
+
+                # First materialize on target device
+                module.to_empty(device=target_device, recurse=False)
+                fqn_path = ".".join(fqn)
+                # Fill weights
+                model.load_weights_to_module(
+                    fqn_path,
+                    weights,
+                )
+                # Quantize weights if applicable
+                if torchao_config and "proj" in fqn_path:
+                    # Note: `None` here is needed to indicate no filter, see
+                    # `apply_torchao_config_to_model` for details.
+                    apply_torchao_config_to_model(module, torchao_config, None)
+
+            # Start calling on root module
+            fill_module(model, [], weights)
+
+        if torchao_config:
+            model.torchao_applied = True
+
+        return model.eval()
+
+
 class DummyModelLoader(BaseModelLoader):
     """Model loader that will set model weights to random values."""
 
@@ -1149,4 +1221,7 @@ def get_model_loader(load_config: LoadConfig) -> BaseModelLoader:
     if load_config.load_format == LoadFormat.GGUF:
         return GGUFModelLoader(load_config)
 
+    if load_config.load_format == LoadFormat.LAYERED:
+        return LayeredModelLoader(load_config)
+
     return DefaultModelLoader(load_config)

python/sglang/srt/models/torch_native_llama.py

@@ -460,7 +460,12 @@ class TorchNativeLlamaForCausalLM(nn.Module):
         params_dict = dict(self.named_parameters())
         return len(params_dict)
 
-    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+    def load_weights_to_module(
+        self,
+        fqn: str,
+        weights: Iterable[Tuple[str, torch.Tensor]],
+    ):
+        """Load weights onto submodule pointed by path `fqn`."""
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
             (".qkv_proj", ".q_proj", "q"),
@@ -469,7 +474,8 @@ class TorchNativeLlamaForCausalLM(nn.Module):
             (".gate_up_proj", ".gate_proj", 0),
             (".gate_up_proj", ".up_proj", 1),
         ]
-        params_dict = dict(self.named_parameters())
+        module = self.get_submodule(fqn)
+        params_dict = dict(module.named_parameters(prefix=fqn, recurse=False))
 
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name or "projector" in name:
@@ -486,7 +492,7 @@ class TorchNativeLlamaForCausalLM(nn.Module):
                     continue
                 name = name.replace(weight_name, param_name)
                 # Skip loading extra bias for GPTQ models.
-                if name.endswith(".bias") and name not in params_dict:
+                if name.endswith(".bias") or name not in params_dict:
                     continue
                 param = params_dict[name]
                 weight_loader = param.weight_loader
@@ -494,12 +500,19 @@ class TorchNativeLlamaForCausalLM(nn.Module):
                 break
             else:
                 # Skip loading extra bias for GPTQ models.
-                if name.endswith(".bias") and name not in params_dict:
+                if name.endswith(".bias") or name not in params_dict:
                     continue
                 param = params_dict[name]
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)
 
+    def load_weights(
+        self,
+        weights: Iterable[Tuple[str, torch.Tensor]],
+    ):
+        """Load weights onto the full model."""
+        self.load_weights_to_module("", weights)
+
 
 class TorchNativePhi3ForCausalLM(TorchNativeLlamaForCausalLM):
     pass

python/sglang/srt/server_args.py

@@ -317,6 +317,7 @@ class ServerArgs:
                 "dummy",
                 "gguf",
                 "bitsandbytes",
+                "layered",
             ],
             help="The format of the model weights to load. "
             '"auto" will try to load the weights in the safetensors format '
@@ -330,7 +331,10 @@ class ServerArgs:
             "which is mainly for profiling."
             '"gguf" will load the weights in the gguf format. '
             '"bitsandbytes" will load the weights using bitsandbytes '
-            "quantization.",
+            "quantization."
+            '"layered" loads weights layer by layer so that one can quantize a '
+            "layer before loading another to make the peak memory envelope "
+            "smaller.",
         )
         parser.add_argument(
             "--trust-remote-code",