Added support for compression on meta device (#376)

* added support for compression on meta device

Signed-off-by: shanjiaz <zsjwpianpian@gmail.com>

* remove breakpoint

Signed-off-by: shanjiaz <zsjwpianpian@gmail.com>

* remove comment

Signed-off-by: shanjiaz <zsjwpianpian@gmail.com>

* address reviewed issues

Signed-off-by: shanjiaz <zsjwpianpian@gmail.com>

* fix style

Signed-off-by: shanjiaz <zsjwpianpian@gmail.com>

* new line

Signed-off-by: shanjiaz <zsjwpianpian@gmail.com>

* Update src/compressed_tensors/compressors/sparse_compressors/sparse_24_bitmask.py

Co-authored-by: Brian Dellabetta <brian-dellabetta@users.noreply.github.com>

* Update src/compressed_tensors/compressors/sparse_compressors/sparse_24_bitmask.py

Co-authored-by: Brian Dellabetta <brian-dellabetta@users.noreply.github.com>

* Update src/compressed_tensors/compressors/sparse_compressors/sparse_24_bitmask.py

Co-authored-by: Brian Dellabetta <brian-dellabetta@users.noreply.github.com>

* Added docstring

Signed-off-by: shanjiaz <zsjwpianpian@gmail.com>

* removed is_meta input

Signed-off-by: shanjiaz <zsjwpianpian@gmail.com>

* added test

Signed-off-by: shanjiaz <zsjwpianpian@gmail.com>

---------

Signed-off-by: shanjiaz <zsjwpianpian@gmail.com>
Co-authored-by: Brian Dellabetta <brian-dellabetta@users.noreply.github.com>

Author: shanjiaz

src/compressed_tensors/compressors/model_compressors/model_compressor.py

@@ -390,9 +390,16 @@ def compress_model(self, model: Module):
         )
 
         for prefix, module in tqdm(model.named_modules(), desc="Compressing model"):
+
             if prefix in module_to_scheme or prefix in sparse_compression_targets:
+                module_device = get_execution_device(module).type
+                is_meta = (module_device == "meta")
+
+                exec_device = "meta" if is_meta else "cpu"
+                onloading_device = "meta" if is_meta else module_device
+
                 # in the future, support compression on same device
-                with align_module_device(module, execution_device="cpu"):
+                with align_module_device(module, execution_device=exec_device):
                     state_dict = module.state_dict(prefix=f"{prefix}.")
 
                 # quantization first
@@ -401,6 +408,7 @@ def compress_model(self, model: Module):
                         state_dict,
                         names_to_scheme=module_to_scheme,
                         show_progress=False,
+                        compression_device=exec_device,
                     )
 
                 # sparsity second
@@ -412,15 +420,14 @@ def compress_model(self, model: Module):
                     )
 
                 # remove any existing parameters
-                exec_device = get_execution_device(module)
                 offload_device = get_offloaded_device(module)
                 for name, _ in list(module.named_parameters()):
                     delete_offload_parameter(module, name)
 
                 # replace with compressed parameters
                 for name, value in state_dict.items():
                     name = name.removeprefix(f"{prefix}.")
-                    value = value.to(exec_device)
+                    value = value.to(onloading_device)
                     param = torch.nn.Parameter(value, requires_grad=False)
                     register_offload_parameter(module, name, param, offload_device)
 

src/compressed_tensors/compressors/quantized_compressors/base.py

@@ -72,6 +72,7 @@ def compress(
         model_state: Dict[str, Tensor],
         names_to_scheme: Dict[str, QuantizationScheme],
         show_progress: bool = False,
+        compression_device: str = "cpu",
         **kwargs,
     ) -> Dict[str, Tensor]:
         """
@@ -85,7 +86,6 @@ def compress(
         """
         uncompressed_names = list(model_state.keys())
         compressed_dict = {}
-        save_device = "cpu"
 
         # compress values
         desc = "Compressing with quantization"
@@ -104,10 +104,10 @@ def compress(
 
                 # is scale does not exist, then weight cannot be compressed
                 if scale is None:
-                    compressed_dict[name] = value.to(save_device)
+                    compressed_dict[name] = value.to(compression_device)
                     continue
 
-                # compress values on cpu (memory movement too expensive)
+                # compress values on meta if loading from meta otherwise on cpu (memory movement too expensive)
                 module_path = prefix[:-1] if prefix.endswith(".") else prefix
                 quant_args = names_to_scheme[module_path].weights
                 compressed_values = self.compress_weight(
@@ -117,12 +117,12 @@ def compress(
                     global_scale=global_scale,
                     g_idx=g_idx,
                     quantization_args=quant_args,
-                    device="cpu",
+                    device=compression_device,
                 )
 
                 # update state dict
                 for key, value in compressed_values.items():
-                    compressed_dict[prefix + key] = value.to(save_device)
+                    compressed_dict[prefix + key] = value.to(compression_device)
 
             else:
                 # omit saving zero points for symmetric or packed quantization
@@ -133,8 +133,7 @@ def compress(
                 # TODO: does this case actually occur?
                 elif name.endswith("g_idx") and torch.any(value <= -1):
                     continue
-
-                compressed_dict[name] = value.to(save_device)
+                compressed_dict[name] = value.to(compression_device)
 
         return compressed_dict
 

src/compressed_tensors/compressors/quantized_compressors/pack_quantized.py

@@ -220,30 +220,34 @@ def pack_to_int32(
     if num_bits < 1:
         raise ValueError(f"num_bits must be at least 1, got {num_bits}")
 
-    # convert to unsigned for packing
+    # Convert to unsigned range for packing, matching quantization offset
     offset = 1 << (num_bits - 1)
     value = (value + offset).to(torch.uint8)
-    value = value.cpu().numpy().astype(np.uint32)
+    device = value.device
+
     pack_factor = 32 // num_bits
 
-    # pad input tensor and initialize packed output
-    packed_size = math.ceil(value.shape[packed_dim] / pack_factor)
-    padding = packed_size * pack_factor - value.shape[packed_dim]
-    value = np.pad(value, pad_width=[(0, 0), (0, padding)], constant_values=0)
+    if packed_dim == 0:
+        value = value.transpose(0, 1)
 
-    # pack values
-    if packed_dim == 1:
-        packed = np.zeros((value.shape[0], packed_size), dtype=np.uint32)
-        for i in range(pack_factor):
-            packed |= value[:, i::pack_factor] << num_bits * i
-    else:
-        packed = np.zeros((packed_size, value.shape[1]), dtype=np.uint32)
-        for i in range(pack_factor):
-            packed |= value[i::pack_factor, :] << num_bits * i
+    rows, cols = value.shape
+    padded_cols = math.ceil(cols / pack_factor) * pack_factor
+    pad_len = padded_cols - cols
+
+    if pad_len > 0:
+        value = torch.nn.functional.pad(value, (0, pad_len))
+
+    num_groups = padded_cols // pack_factor
+
+    # Use int32 here
+    reshaped = value.view(rows, num_groups, pack_factor).to(torch.int32)
+    bit_shifts = torch.arange(pack_factor, device=device, dtype=torch.int32) * num_bits
+    packed = (reshaped << bit_shifts).sum(dim=2, dtype=torch.int32)
+
+    if packed_dim == 0:
+        packed = packed.transpose(0, 1)
 
-    # convert back to signed and torch
-    packed = np.ascontiguousarray(packed).view(np.int32)
-    return torch.from_numpy(packed)
+    return packed
 
 
 def unpack_from_int32(

src/compressed_tensors/compressors/sparse_compressors/sparse_24_bitmask.py

@@ -56,8 +56,10 @@ def compress_weight(self, name, value):
         bitmask_tensor = Sparse24BitMaskTensor.from_dense(
             value, self.config.sparsity_structure
         )
-        bitmask_dict = bitmask_tensor.dict(name_prefix=name, device="cpu")
-        return bitmask_dict
+        return bitmask_tensor.dict(
+            name_prefix=name,
+            device="meta" if value.is_meta else "cpu",
+        )
 
     def decompress_weight(self, weight_data):
         data = Sparse24BitMaskTensor.from_compressed_data(**weight_data)
@@ -90,9 +92,14 @@ def from_dense(
         :return: instantiated compressed tensor
         """
         shape = list(tensor.shape)
-        compressed, bitmask = sparse24_bitmask_compress(
-            tensor.cpu(), sparsity_structure=sparsity_structure
-        )
+        if tensor.is_meta:
+            compressed, bitmask = sparse24_bitmask_compress(
+                tensor, sparsity_structure=sparsity_structure
+            )
+        else:
+            compressed, bitmask = sparse24_bitmask_compress(
+                tensor.cpu(), sparsity_structure=sparsity_structure
+            )
         return Sparse24BitMaskTensor(
             shape=shape,
             compressed=compressed,
@@ -169,6 +176,13 @@ def sparse24_bitmask_compress(
         SparsityStructure(sparsity_structure) == SparsityStructure.TWO_FOUR
     ), "Only 2:4 sparsity is supported"
 
+    if tensor.is_meta:
+        num_rows, num_cols = tensor.shape
+        compressed_values = torch.empty((num_rows, num_cols // 2), dtype=tensor.dtype, device="meta")
+        packed_cols = (num_cols + 7) // 8
+        bitmasks_packed = torch.empty((num_rows, packed_cols), dtype=torch.uint8, device="meta")
+        return compressed_values, bitmasks_packed
+
     bytemasks = get_24_bytemasks(tensor=tensor)
 
     if tensor.dtype == FP8_DTYPE:

tests/test_compressors/model_compressors/test_model_compressor.py

@@ -412,6 +412,66 @@ def test_compress_model(model_stub, q_format, s_config, tmpdir):
         assert torch.all(compressed[key] == true_compressed[key]), f"{key}"
 
 
+@pytest.mark.parametrize(
+    "model_stub,q_format,s_config",
+    [
+        (
+            "nm-testing/llama2.c-stories42M-gsm8k-quantized-only-uncompressed",
+            "float-quantized",
+            None,
+        ),
+        (
+            "nm-testing/llama2.c-stories42M-gsm8k-sparse-only-uncompressed",
+            None,
+            "sparse-24-bitmask",
+        ),
+        (
+            "nm-testing/llama2.c-stories42M-gsm8k-stacked-uncompressed",
+            "float-quantized",
+            "sparse-24-bitmask",
+        ),
+        (
+            "nm-testing/llama2.c-stories15M-ultrachat-mixed-uncompressed",
+            "pack-quantized",
+            None,
+        ),
+    ],
+)
+def test_compress_model_meta(model_stub, q_format, s_config):
+    # Load model on CPU to get expected compressed state_dict
+    cpu_model = AutoModelForCausalLM.from_pretrained(
+        model_stub, torch_dtype=torch.float32
+    )
+    reference_compressor = ModelCompressor.from_pretrained_model(
+        cpu_model, s_config, q_format
+    )
+    # Only stores dtype because meta model does not store values
+    expected = {
+        k: v.dtype
+        for k, v in reference_compressor.compress(cpu_model).items()
+    }
+
+    # Load model on meta device
+    meta_model = AutoModelForCausalLM.from_pretrained(
+        model_stub,
+        torch_dtype=torch.float32,
+        low_cpu_mem_usage=True,
+    )
+    for module in meta_model.modules():
+        if hasattr(module, "to_empty"):
+            module.to_empty(device="meta")
+
+    # Compress in-place on meta model
+    compressor = ModelCompressor.from_pretrained_model(meta_model, s_config, q_format)
+    compressor.compress_model(meta_model)
+
+    # Compare keys and dtypes
+    compressed = dict(meta_model.state_dict())
+    assert set(compressed.keys()) == set(expected.keys())
+    for key, dtype in expected.items():
+        assert compressed[key].dtype == dtype, f"{key} has incorrect dtype"
+
+
 @pytest.mark.parametrize(
     "model_stub,comp_stub",
     [