[ Misc ] Refactor w8a8 to use process_weights_after_load (Simplify Weight Loading)

tests/quantization/test_compressed_tensors.py

@@ -11,14 +11,16 @@
     CompressedTensorsLinearMethod, CompressedTensorsW4A16Sparse24,
     CompressedTensorsW8A8DynamicToken, CompressedTensorsW8A8StaticTensor,
     CompressedTensorsWNA16)
+from vllm.model_executor.layers.quantization.compressed_tensors.utils import (
+    QuantizationType)
 
 
 @pytest.mark.parametrize("model_args", [
-    ("nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change", "tensor"),
-    ("nm-testing/tinyllama-oneshot-w8-channel-a8-tensor", "channel"),
+    ("nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change", "tensor", QuantizationType.INT),
+    ("nm-testing/tinyllama-oneshot-w8-channel-a8-tensor", "channel", QuantizationType.INT),
 ])
 def test_compressed_tensors_w8a8_static_setup(vllm_runner, model_args):
-    model_path, strategy = model_args
+    model_path, strategy, quant_type = model_args
     with vllm_runner(model_path, enforce_eager=True) as llm:
         model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model  # noqa: E501
         layer = model.model.layers[0]
@@ -34,17 +36,21 @@ def test_compressed_tensors_w8a8_static_setup(vllm_runner, model_args):
                           CompressedTensorsLinearMethod)
         assert isinstance(down_proj.quant_method,
                           CompressedTensorsLinearMethod)
-
         assert isinstance(qkv_proj.scheme, CompressedTensorsW8A8StaticTensor)
 
         assert qkv_proj.scheme.strategy == strategy
-        assert qkv_proj.weight.dtype is torch.int8
-        assert o_proj.weight.dtype is torch.int8
-        assert gate_up_proj.weight.dtype is torch.int8
+        expected_type = (torch.int8 
+                         if quant_type == QuantizationType.INT
+                         else torch.float8_e4m3fn)
+
+        assert qkv_proj.weight.dtype is expected_type
+        assert o_proj.weight.dtype is expected_type
+        assert gate_up_proj.weight.dtype is expected_type
 
         if qkv_proj.scheme.strategy == "tensor":
-            assert qkv_proj.weight_scale.shard_splitter is not None
-            assert qkv_proj.weight_scale.logical_widths is not None
+            breakpoint()
+            assert qkv_proj.weight_scale.is_per_tensor_scale
+        assert qkv_proj.weight_scale.dtype is torch.float32
         assert qkv_proj.input_scale.dtype is torch.float32
 
 

vllm/model_executor/layers/linear.py

@@ -41,6 +41,29 @@ def adjust_bitsandbytes_shard(param: Parameter,
     return quantized_size, quantized_offset
 
 
+def adjust_per_tensor_scale_shard(param, loaded_weight, shard_id):
+    """For fused modules (QKV and MLP) we have an array of length
+    N that holds 1 scale for each "logical" matrix. So the param
+    is an array of length N. The loaded_weight corresponds to 
+    one of the shards on disk. Here, we slice the param based on 
+    the shard_id for loading.
+    """
+    qkv_idxs = {"q": 0, "k": 1, "v": 2}
+
+    if isinstance(shard_id, str):
+        shard_id = qkv_idxs[shard_id]
+    elif not isinstance(shard_id, int):
+        raise ValueError(f"Unknown Shard Id {shard_id}")
+
+    # AutoFP8 scales do not have a shape
+    # compressed-tensors scales do have a shape
+    if len(loaded_weight.shape) != 0:
+        assert loaded_weight.shape[0] == 1
+        loaded_weight = loaded_weight[0]
+
+    return param[shard_id], loaded_weight
+
+
 class LinearMethodBase(QuantizeMethodBase):
     """Base class for different (maybe quantized) linear methods."""
 
@@ -269,10 +292,6 @@ def __init__(self,
             self.register_parameter("bias", None)
 
     def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
-        # Special case for Fp8 scales.
-        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
-                                           None)
-
         tp_rank = get_tensor_model_parallel_rank()
         output_dim = getattr(param, "output_dim", None)
         param_data = param.data
@@ -281,11 +300,11 @@ def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
             start_idx = tp_rank * shard_size
             loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                  shard_size)
-        # Special case for Fp8 scales.
-        elif fp8_scales_shard_indexer is not None:
-            param_data, loaded_weight = fp8_scales_shard_indexer(param_data,
-                                                                 loaded_weight,
-                                                                 shard_id=0)
+
+        # Special case for loading scales off disk, which often
+        # do not have a shape.
+        if len(loaded_weight.shape) == 0:
+            loaded_weight = loaded_weight.reshape(1)
 
         assert param_data.shape == loaded_weight.shape
         param_data.copy_(loaded_weight)
@@ -362,29 +381,8 @@ def weight_loader(self,
         output_dim = getattr(param, "output_dim", None)
         # Special case for AQLM codebooks.
         is_metadata = getattr(param, "is_metadata", False)
-
-        param_shard_splitter = getattr(param, "shard_splitter", None)
-
-        if output_dim is not None and param_shard_splitter is not None:
-            raise NotImplementedError(
-                "We do not currently support output_dim != None and "
-                "shard_splitter != None for a parameter. Please open an issue."
-            )
-        # If a parameter has defined a shard_splitter to be used for
-        # the weight, it should be applied before the weight is
-        # loaded/copied to the parameter. The shard_splitter applies
-        # logic by using the loaded_shard_id to ensure that the loaded
-        # param is loaded to the correct location
-        # within the parameter defined by the linear method.
-        if loaded_shard_id is None and param_shard_splitter is not None:
-            raise NotImplementedError(
-                "We do not currently support loaded_shard_id == None and "
-                "shard_splitter != None for a parameter. Please open an issue."
-            )
-
-        # Special case for Fp8 scales.
-        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
-                                           None)
+        # Special case for per-tensor scale in fused case.
+        is_per_tensor_scale = getattr(param, "is_per_tensor_scale", False)
 
         if loaded_shard_id is None:
             # Loaded weight is already packed.
@@ -449,15 +447,9 @@ def weight_loader(self,
             shard_offset = loaded_shard_id * shard_size
             param_data = param_data.narrow(0, shard_offset, shard_size)
 
-        # If a param_shard_splitter is defined by the LinearMethod, use it.
-        elif param_shard_splitter is not None:
-            logical_widths = getattr(param, "logical_widths", None)
-            param_data, loaded_weight = param_shard_splitter(
-                param_data, loaded_weight, loaded_shard_id, logical_widths)
-
-        # Special case for Fp8 scales.
-        elif fp8_scales_shard_indexer is not None:
-            param_data, loaded_weight = fp8_scales_shard_indexer(
+        # Special case for per-tensor scales in fused case.
+        elif is_per_tensor_scale:
+            param_data, loaded_weight = adjust_per_tensor_scale_shard(
                 param_data, loaded_weight, loaded_shard_id)
 
         else:
@@ -547,28 +539,8 @@ def weight_loader(self,
         # Special case for AQLM codebooks.
         is_metadata = getattr(param, "is_metadata", False)
 
-        param_shard_splitter = getattr(param, "shard_splitter", None)
-
-        if output_dim is not None and param_shard_splitter is not None:
-            raise NotImplementedError(
-                "We do not currently support output_dim != None and "
-                "shard_splitter != None for a parameter. Please open an issue."
-            )
-        # If a parameter has defined a shard_splitter to be used for
-        # the weight, it should be applied before the weight is
-        # loaded/copied to the parameter. The shard_splitter applies
-        # logic by using the loaded_shard_id to ensure that the loaded
-        # param is loaded to the correct location
-        # within the parameter defined by the linear method.
-        if loaded_shard_id is None and param_shard_splitter is not None:
-            raise NotImplementedError(
-                "We do not currently support loaded_shard_id == None and "
-                "shard_splitter != None for a parameter. Please open an issue."
-            )
-
-        # Special case for Fp8 scales.
-        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
-                                           None)
+        # Special case for per-tensor scales in fused case.
+        is_per_tensor_scale = getattr(param, "is_per_tensor_scale", False)
 
         if loaded_shard_id is None:
             # Loaded weight is already packed.
@@ -661,15 +633,9 @@ def weight_loader(self,
             shard_index = ["q", "k", "v"].index(loaded_shard_id)
             param_data = param_data.narrow(0, shard_index * shard_size,
                                            shard_size)
-        # If a param_shard_splitter is defined by the LinearMethod, use it.
-        elif param_shard_splitter is not None:
-            logical_widths = getattr(param, "logical_widths", None)
-            param_data, loaded_weight = param_shard_splitter(
-                param_data, loaded_weight, loaded_shard_id, logical_widths)
-
-        # Special case for Fp8 scales.
-        elif fp8_scales_shard_indexer is not None:
-            param_data, loaded_weight = fp8_scales_shard_indexer(
+        # Special case for per-tensor scales in fused case.
+        elif is_per_tensor_scale:
+            param_data, loaded_weight = adjust_per_tensor_scale_shard(
                 param_data, loaded_weight, loaded_shard_id)
         else:
             ignore_warning = getattr(param, "ignore_warning", False)
@@ -751,10 +717,6 @@ def __init__(self,
             self.register_parameter("bias", None)
 
     def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
-        # Special case for Fp8 scales.
-        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
-                                           None)
-
         tp_rank = get_tensor_model_parallel_rank()
         input_dim = getattr(param, "input_dim", None)
         param_data = param.data
@@ -764,13 +726,7 @@ def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
             loaded_weight = loaded_weight.narrow(input_dim, start_idx,
                                                  shard_size)
 
-        # Special case for Fp8 scales.
-        elif fp8_scales_shard_indexer is not None:
-            param_data, loaded_weight = fp8_scales_shard_indexer(param_data,
-                                                                 loaded_weight,
-                                                                 shard_id=0)
-
-        if fp8_scales_shard_indexer is None and len(loaded_weight.shape) == 0:
+        if len(loaded_weight.shape) == 0:
             loaded_weight = loaded_weight.reshape(1)
 
         assert param_data.shape == loaded_weight.shape

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py

@@ -177,6 +177,9 @@ class CompressedTensorsLinearMethod(LinearMethodBase):
     def __init__(self, quantization_config: CompressedTensorsConfig):
         self.quantization_config = quantization_config
 
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        return layer.scheme.process_weights_after_loading(layer)
+
     def create_weights(self, layer: torch.nn.Module,
                        input_size_per_partition: int,
                        output_partition_sizes: List[int], input_size: int,

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_scheme.py

@@ -31,3 +31,11 @@ def apply_weights(self, layer: torch.nn.Module, x: torch.Tensor):
 
         """
         raise NotImplementedError
+
+    @abstractmethod
+    def process_weights_after_loading(self, layer: torch.nn.Module):
+        """
+        Called after weight loading is complete for any cleanup that
+        needs to occur.
+        """
+        raise NotImplementedError

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_unquantized.py

@@ -18,6 +18,9 @@ class CompressedTensorsUnquantized(CompressedTensorsScheme):
     in a linear transformation.
     """
 
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        pass
+
     def create_weights(self, layer: torch.nn.Module,
                        output_partition_sizes: List[int],
                        input_size_per_partition: int,

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a16_24.py

@@ -29,6 +29,9 @@ def __init__(self,
             raise ValueError(
                 "group_size must be given when using strategy group")
 
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        pass
+
     def create_weights(self, layer: torch.nn.Module, input_size: int,
                        output_partition_sizes: List[int],
                        input_size_per_partition: int,

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8.py

@@ -15,70 +15,63 @@ class CompressedTensorsW8A8(CompressedTensorsScheme):
     def __init__(self, strategy: str):
         self.strategy = strategy
 
-    def _shard_id_as_int(self, shard_id: Union[str, int]) -> int:
-        if isinstance(shard_id, int):
-            return shard_id
-
-        assert isinstance(shard_id, str)
-        qkv_idxs = {"q": 0, "k": 1, "v": 2}
-        assert shard_id in qkv_idxs
-        return qkv_idxs[shard_id]
-
-    def scales_shard_splitter(
-            self, param: torch.Tensor, loaded_weight: torch.Tensor,
-            shard_id: Union[str, int],
-            logical_widths: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-        shard_id = self._shard_id_as_int(shard_id)
-        offset = sum(logical_widths[:shard_id])
-        size = logical_widths[shard_id]
-        # update loaded weight with copies for broadcast.
-        loaded_weight = loaded_weight.repeat(size)
-        return param[offset:offset + size], loaded_weight
+    # Cutlass kernels support only per-tensor and per-channel cases.
+    # So if we have a fused module (QKV, MLP) with per tensor scales (thus N
+    # scales being passed to the kernel), we convert to the per-channel case.
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        if (self.strategy == QuantizationStrategy.TENSOR
+                and len(self.logical_widths) > 1):
+
+            # Load the N per-tensor scales into the channelwise buffer.
+            weight_scale_channel = torch.empty(
+                (sum(self.logical_widths), 1),
+                dtype=torch.float32,
+                device=layer.weight_scale.device)
+            start = 0
+            for idx, logical_width in enumerate(self.logical_widths):
+                end = start + logical_width
+                weight_scale_channel[start:end, :] = layer.weight_scale[idx]
+                start = end
+
+            layer.weight_scale = Parameter(weight_scale_channel,
+                                           requires_grad=False)
 
     def create_weights(self, layer: torch.nn.Module,
                        output_partition_sizes: List[int],
                        input_size_per_partition: int,
                        params_dtype: torch.dtype, weight_loader: Callable,
                        **kwargs):
+        self.logical_widths = output_partition_sizes
 
-        is_tensor_partitioned = len(output_partition_sizes) != 1
-        weight_scale_dim = sum(output_partition_sizes) if (
-            is_tensor_partitioned
-            or self.strategy == QuantizationStrategy.CHANNEL) else 1
-
-        shape: Union[Tuple[int], Tuple[int, int]] = (weight_scale_dim, )
+        # WEIGHT SCALE
+        shape: Union[Tuple[int], Tuple[int, int]]
         if self.strategy == QuantizationStrategy.CHANNEL:
-            shape = (weight_scale_dim, 1)
+            shape = (sum(self.logical_widths), 1)
+        else:
+            shape = (len(self.logical_widths), )
 
         weight_scale = Parameter(torch.empty(*shape, dtype=torch.float32),
                                  requires_grad=False)
-
         layer.register_parameter("weight_scale", weight_scale)
-        set_weight_attrs(weight_scale, {"weight_loader": weight_loader})
+        if self.strategy == QuantizationStrategy.CHANNEL:
+            set_weight_attrs(weight_scale, {
+                "weight_loader": weight_loader,
+                "output_dim": 0,
+            })
+        else:
+            set_weight_attrs(weight_scale, {
+                "weight_loader": weight_loader,
+                "is_per_tensor_scale": True,
+            })
 
+        # WEIGHT
         weight = Parameter(torch.empty(sum(output_partition_sizes),
                                        input_size_per_partition,
                                        dtype=torch.int8),
                            requires_grad=False)
-
         layer.register_parameter("weight", weight)
-        set_weight_attrs(
-            weight, {
-                "input_dim": 1,
-                "output_dim": 0,
-                "weight_loader": weight_loader,
-                "logical_widths": output_partition_sizes
-            })
-
-        # Don't need a shard_splitter for channel-wise quantization
-        # Use the default loading method
-        if self.strategy == QuantizationStrategy.CHANNEL:
-            set_weight_attrs(weight_scale, {
-                "output_dim": 0,
-            })
-        else:
-            set_weight_attrs(
-                weight_scale, {
-                    "logical_widths": output_partition_sizes,
-                    "shard_splitter": self.scales_shard_splitter,
-                })
+        set_weight_attrs(weight, {
+            "input_dim": 1,
+            "output_dim": 0,
+            "weight_loader": weight_loader,
+        })

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16.py

@@ -29,6 +29,9 @@ def __init__(self,
             raise ValueError(
                 "group_size must be given when using strategy group")
 
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        pass
+
     def create_weights(self, layer: torch.nn.Module, input_size: int,
                        output_partition_sizes: List[int],
                        input_size_per_partition: int,