Update coreml codebook (#2648)

* Update CoreML codebook APIs

* up

* up

* up

Author: metascroy

test/prototype/test_codebook_coreml.py

@@ -14,7 +14,6 @@
 )
 from torchao.quantization import quantize_
 from torchao.quantization.utils import compute_error
-from torchao.testing.utils import skip_if_no_cuda
 from torchao.utils import TORCH_VERSION_AT_LEAST_2_6, is_package_at_least
 
 
@@ -36,7 +35,7 @@ def test_choose_qparams_codebook(self):
             self.block_size,
         )
         group_size = self.block_size[-1]
-        self.assertEqual(codebook.shape, (256 // group_size, 2**self.nbits, 1))
+        self.assertEqual(codebook.shape, (1, 256 // group_size, 2**self.nbits, 1))
         self.assertEqual(wq.shape, (100, 256))
 
         self.assertFalse(torch.isnan(codebook).any())
@@ -76,7 +75,6 @@ def test_quantize_api(self):
         )
         assert type(m[0].weight) == CodebookQuantizedTensor
 
-    @skip_if_no_cuda()
     @unittest.skipIf(not TORCH_VERSION_AT_LEAST_2_6, "requires 2.6+.")
     def test_export(self):
         m = torch.nn.Sequential(torch.nn.Linear(128, 64)).to(torch.float32)

torchao/prototype/quantization/codebook_coreml/api.py

@@ -42,13 +42,12 @@ def _codebook_weight_only_transform(
         raise ImportError("Requires coremltools >= 8.3.0")
 
     dtype = config.dtype
-    block_size = config.block_size
     weight = module.weight
 
     quantized_weight = CodebookQuantizedTensor.from_float(
         weight,
         dtype,
-        block_size,
+        config.block_size,
     )
     module.weight = torch.nn.Parameter(quantized_weight, requires_grad=False)
     return module

torchao/prototype/quantization/codebook_coreml/codebook_ops.py

@@ -34,11 +34,12 @@ def choose_qparams_and_quantize_codebook_coreml(
     Args:
         input_tensor (torch.Tensor): The input tensor to be quantized.
         code_dtype (torch.dtype): The dtype for the codes. [torch.uint1, ..., torch.uint8]
-        block_size (List[int]): the size for how many elements of last dimension of input_tensor
-          belong to the same group and should share the same lookup table. let's say original
-          shape is (N, K), and block_size of (N, group_size) or (-1, group_size),
-          then the slice of (N, group_size) elements should use the same lookup
-          table, and there will be (K // group_size) lookup tables
+        block_size (List[int]): block sizes for how many elements in each dimension share
+           the same lookup table (len(block_size) == input_tensor.dim())
+           Each dimension of input_tensor must be divisible by the corresponding element of block_size
+           Look up tables are indexed by {(di // bi) for i in input_tensor.dim()}
+           For example, if the input tensor has shape (N, K), and block_size is (N, group_size), this means
+           there is a lookup table for group_size columns, i.e., (K // group_size) total look up tables
         force_kmeans1d (bool): Use kmeans1d regardless of number of weights
         cluster_dim (int): this means the size of the vector for vector lookup table quantization
           e.g. when cluster_dim is 4, instead of quantizing each scalar value one by one, we quantize
@@ -48,43 +49,45 @@ def choose_qparams_and_quantize_codebook_coreml(
 
     Returns:
         Tuple[torch.Tensor, torch.Tensor]  The codebook (lookup table) Tensor and the quantized Tensor (codes, torch.uint8)
+        The LUT table has dimension (g0, .., g(N-1), 2**nbits, vec_dim), where:
+         * The first N dimensions index over the different tables (gi = input_tensor.shape[i] // block_size[i] in each dimension)
+         * The N + 1 dimension indexes over the nbit indices (2 ** nbits)
+         * The N + 2 dimension indexes over the look up values (shape = 1 for scalar)
     """
     assert code_dtype in list(_SUB_BYTE_UINT_BOUNDS.keys()) + [torch.uint8]
-    assert len(block_size) == input_tensor.ndim
+    nbits = _DTYPE_TO_BIT_WIDTH[code_dtype]
+    assert nbits >= 1 and nbits <= 8, f"nbits must be in [1, 8], got {nbits}"
+
+    assert len(block_size) == input_tensor.dim()
     block_size = block_size.copy()
-    for i in range(input_tensor.ndim - 1):
-        assert block_size[i] == -1 or block_size[i] == input_tensor.shape[i], (
-            f"{block_size} not supported"
+    for i in range(len(block_size)):
+        if block_size[i] == -1:
+            block_size[i] = input_tensor.shape[i]
+        assert block_size[i] >= 1 and input_tensor.shape[i] % block_size[i] == 0, (
+            "block_size[i] must divide input_tensor.shape[i]"
         )
 
-    group_size = block_size[-1]
-    if group_size == -1:
-        group_size = input_tensor.shape[-1]
-
-    assert input_tensor.shape[-1] % group_size == 0
-    assert input_tensor.ndim == 2
+    assert input_tensor.dim() == 2, "Currently only rank 2 tensors are supported"
+    assert block_size[0] == input_tensor.shape[0], (
+        "Currently only support per-grouped channel granularity"
+    )
     assert cluster_dim == 1, (
         f"only cluster_dim == 1 is supported right now, got {cluster_dim}"
     )
 
+    num_lut = input_tensor.shape[1] // block_size[1]
+    group_size = block_size[1]
+
     # for converting to numpy
     input_tensor = input_tensor.detach()
-    # (N, K)
     original_shape = input_tensor.shape
-    # (K // group_size)
-    num_lut = input_tensor.shape[1] // group_size
 
     # reshape to (N, K // group_size, group_size)
     input_tensor = input_tensor.reshape(input_tensor.shape[0], num_lut, group_size)
     from coremltools.models.neural_network.quantization_utils import (
         _get_kmeans_lookup_table_and_weight,
     )
 
-    nbits = _DTYPE_TO_BIT_WIDTH[code_dtype]
-    if nbits > 8:
-        print(f"Requested nbits: {nbits}, rewriting to 8 bits to reduce the size")
-        nbits = 8
-
     res_lut = []
     # each res_w[:, i, :] will use the same lookup table
     # res_w: (N, K // group_size, group_size)
@@ -102,6 +105,13 @@ def choose_qparams_and_quantize_codebook_coreml(
     # res_lut: (K // group_size, 2 ** nbits)
     res_lut = torch.stack(res_lut, dim=0)
 
+    # The final LUT should have dimension equal to input_tensor.dim() + 2
+    # The first input_tensor.dim() dimensions index over the tables,
+    # input_tensor.dim() + 1 indexes over the nbit indices
+    # input_tensor.dim() + 2 are the look up values (shape = 1 for scalar)
+    # res_lut: (N, K // group_size, 2 ** nbits, group_size)
+    res_lut = res_lut.reshape(1, num_lut, 2**nbits, 1)
+
     # reshape back to (N, K)
     res_w = res_w.reshape(*original_shape)
 
@@ -112,7 +122,7 @@ def choose_qparams_and_quantize_codebook_coreml(
 def dequantize_codebook(
     codes: torch.Tensor,
     codebook: torch.Tensor,
-    code_dtype: torch.dtype,
+    nbits: int,
     block_size: List[int],
     output_dtype: torch.dtype = torch.float32,
 ) -> torch.Tensor:
@@ -121,13 +131,14 @@ def dequantize_codebook(
 
     Args:
         codes (torch.Tensor): Indices of codebook entries for each element
-                              shape (N, K) for scalar quantization
-        codebook (torch.Tensor): Codebook tensor used for quantization,
-                                 shape (K // group_size, 2 ** nbits) where K is the dim 1 shape of input
-        code_dtype (torch.dtype): The logical dtype for the codes, [torch.uint1, ..., torch.uint8]
-         Note that codes is stored in torch.uint8, this is just addtional information for dequantize op
-        block_size (List[int]): a slice of elements with shape block_size will share the same lookup table
-            only support (-1, ..., group_size) right now (all preceding dimensions has to match input)
+            General shape: (d0, d1, d2, ..., dN)
+            Simple example shape: (N, K)
+        codebook (torch.Tensor): Codebook tensor used for quantization
+            General shape: (d0 // block_size[0], ..., dN // block_size[N], 2**nbits, vec_dim), where vec_dim = 1 for scalar look up values
+            Simple example shape: (1, group_size, 2 ** nbits, 1) for scalar look up values, with 1 table per group_size columns
+        nbits: int: number of bits for the quantization
+        block_size (List[int]): a slice of elements with shape block_size will share the same lookup table.
+            If block_size[i] == -1, then the entire dimension is used.
         output_dtype (torch.dtype): dtype for the output tensor.
 
     Returns:
@@ -140,37 +151,67 @@ def dequantize_codebook(
         torch.bfloat16,
     ], f"Unsupported output dtype: {output_dtype}"
 
-    assert code_dtype in list(_SUB_BYTE_UINT_BOUNDS.keys()) + [torch.uint8]
+    assert nbits >= 1 and nbits <= 8, f"nbits must be in [1, 8], got {nbits}"
 
-    assert len(block_size) == codes.ndim
+    assert len(block_size) == codes.dim()
     block_size = block_size.copy()
-    for i in range(codes.ndim - 1):
-        assert block_size[i] == -1 or block_size[i] == codes.shape[i], (
-            f"{block_size} not supported"
+    for i in range(len(block_size)):
+        if block_size[i] == -1:
+            block_size[i] = codes.shape[i]
+        assert block_size[i] >= 1 and codes.shape[i] % block_size[i] == 0, (
+            "block_size[i] must divide codes.shape[i]"
         )
 
-    group_size = block_size[-1]
-    if group_size == -1:
-        group_size = codes.shape[-1]
+    assert codebook.dim() == codes.dim() + 2
+    codebook_shape = codebook.shape
+    vec_dim = codebook_shape[-1]
+    quant_levels = 2**nbits
 
-    assert codes.shape[-1] % group_size == 0
-    K = codes.shape[-1]
-    num_lut = K // group_size
-    # (N, K)
-    original_shape = codes.shape
+    # Check that last two dimensions of codebook are [quant_levels, vec_dim]
+    assert codebook_shape[-2] == quant_levels, "Codebook shape mismatch with nbits"
 
-    # reshape to (N, num_lut, group_size)
-    codes = codes.reshape(codes.shape[0], num_lut, group_size)
-    dequant = torch.zeros_like(codes, dtype=output_dtype)
+    # Compute shape of lookup group indices from codes shape and block size
+    code_shape = codes.shape
+    ndim = codes.ndim
+    assert len(block_size) == ndim, "block_size must match dimensionality of codes"
 
-    # do lookup for each lookup table
-    # dequant shape: (N, num_lut, group_size)
-    # codebook shape: (num_lut, 2 ** nbits)
-    # codes shape: (N, num_lut, group_size)
-    for i in range(num_lut):
-        # dequant[:, i, :]: (N, group_size)
-        # using squeeze to remove the training dim 1s after the lookup
-        dequant[:, i, :] = codebook[i][codes[:, i, :]].squeeze()
+    # Compute which codebook slice to use for each element
+    group_indices = []
+    for i in range(ndim):
+        assert block_size[i] >= 1 and code_shape[i] % block_size[i] == 0, (
+            f"dimension {code_shape[i]} not divisible by block size {block_size[i]}"
+        )
 
-    dequant = dequant.reshape(*original_shape)
-    return dequant.to(output_dtype)
+        # Index of block
+        idx = (
+            torch.arange(code_shape[i], device=codes.device) // block_size[i]
+        )  # shape (di,)
+
+        # Reshape idx to broadcast along all other dims
+        shape = [1] * ndim
+        shape[i] = code_shape[i]
+        idx = idx.view(*shape)  # shape (1, ..., 1, di, 1, ..., 1)
+        idx = idx.expand(code_shape)  # shape (d0, ..., dN)
+        group_indices.append(idx)
+
+    # Stack the broadcasted group indices
+    # group_index_tensor at (i0, i1, ..., iN) is the gives the group indices (g0, ..., gN)
+    # for the element at (i0, i1, ..., iN) in the original code
+    # If code.shape = (d1, d2, d3), then group_index_tensor.shape = (d1, d2, d3, 3)
+    group_index_tensor = torch.stack(
+        group_indices, dim=-1
+    )  # shape (d0, d1, ..., dN, ndim)
+
+    # Flatten everything to index efficiently
+    flat_codes = codes.reshape(-1)  # shape (numel,)
+    flat_groups = group_index_tensor.reshape(-1, ndim)  # (numel, ndim)
+
+    # Compute dequantized values via indexing
+    # index into codebook with (*group_index, code_index, :)
+    gathered = codebook[(*flat_groups.T, flat_codes)]  # shape (numel, vec_dim)
+    dequant = gathered.reshape(*code_shape, vec_dim)
+
+    if vec_dim == 1:
+        dequant = dequant.squeeze(-1)
+
+    return dequant.to(dtype=output_dtype)

torchao/prototype/quantization/codebook_coreml/codebook_quantized_tensor.py

@@ -12,6 +12,9 @@
     choose_qparams_and_quantize_codebook_coreml,
     dequantize_codebook,
 )
+from torchao.quantization.quant_primitives import (
+    _DTYPE_TO_BIT_WIDTH,
+)
 from torchao.utils import TorchAOBaseTensor
 
 aten = torch.ops.aten
@@ -95,7 +98,7 @@ def dequantize(self, output_dtype: Optional[torch.dtype] = None) -> torch.Tensor
         return dequantize_codebook(
             codes,
             self.codebook,
-            self.code_dtype,
+            _DTYPE_TO_BIT_WIDTH[self.code_dtype],
             self.block_size,
             output_dtype=output_dtype,
         )
@@ -174,6 +177,17 @@ def _(func, types, args, kwargs):
     return func(input_tensor, weight_tensor, bias)
 
 
+@implements([torch.nn.functional.embedding, aten.embedding.default])
+def _(func, types, args, kwargs):
+    assert len(args) == 2
+    indices, weight_tensor = (
+        args[0],
+        args[1],
+    )
+    weight_tensor = weight_tensor.dequantize()
+    return func(indices, weight_tensor, **kwargs)
+
+
 @implements([aten.detach.default, aten.alias.default])
 def _(func, types, args, kwargs):
     return return_and_correct_aliasing(