vllm-cpu 0.11.0.post2__cp312-cp312-manylinux_2_17_x86_64.whl

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

@@ -0,0 +1,238 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from collections.abc import Generator
+from itertools import accumulate
+from typing import Callable, Optional
+
+import torch
+from compressed_tensors.transform import (TransformArgs, TransformConfig,
+                                          TransformLocation, TransformScheme)
+from compressed_tensors.utils import is_match
+
+from vllm.model_executor.layers.linear import (WEIGHT_LOADER_V2_SUPPORTED,
+                                               LinearMethodBase,
+                                               QKVCrossParallelLinear)
+from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import (  # noqa: E501
+    CompressedTensorsScheme)
+from vllm.model_executor.layers.quantization.compressed_tensors.transform.module import (  # noqa: E501
+    HadamardTransform)
+from vllm.model_executor.layers.quantization.compressed_tensors.transform.utils import (  # noqa: E501
+    TransformTuple)
+
+
+class CompressedTensorsLinearTransformMethod(LinearMethodBase):
+    """
+    Wraps `CompressedTensorsLinearMethod` or `UnquantizedLinearMethod` and adds
+    input and output transforms to either side of the original apply method
+    """
+
+    @classmethod
+    def from_schemes(
+        cls,
+        quant_method: LinearMethodBase,
+        quant_scheme: Optional[CompressedTensorsScheme],
+        input_tfms: dict[int, TransformTuple],
+        output_tfms: dict[int, TransformTuple],
+    ) -> "CompressedTensorsLinearTransformMethod":
+        from vllm.model_executor.layers.quantization.compressed_tensors.transform.schemes.linear_qutlass_nvfp4 import (  # noqa: E501
+            QutlassNvFP4LinearMethod, is_qutlass_fp4_scheme)
+
+        assert input_tfms or output_tfms
+
+        if is_qutlass_fp4_scheme(quant_scheme, input_tfms):
+            return QutlassNvFP4LinearMethod(quant_method, input_tfms,
+                                            output_tfms)
+
+        # hadacore or dense gemm is selected by Transform module
+
+        return cls(quant_method, input_tfms, output_tfms)
+
+    def __init__(self, quant_method: LinearMethodBase,
+                 input_tfms: dict[int, TransformTuple],
+                 output_tfms: dict[int, TransformTuple]):
+        self.quant_method = quant_method
+        self.input_tfms = input_tfms
+        self.output_tfms = output_tfms
+
+        self.input_transform: Optional[HadamardTransform] = None
+        self.output_transform: Optional[HadamardTransform] = None
+
+    def create_weights(self, layer: torch.nn.Module,
+                       input_size_per_partition: int,
+                       output_partition_sizes: list[int], input_size: int,
+                       output_size: int, params_dtype: torch.dtype,
+                       **extra_weight_attrs):
+
+        # get weight loader for transforms
+        weight_loader: Callable = extra_weight_attrs.get(
+            "weight_loader")  # type: ignore[assignment]
+
+        # HACK: UnquantizedLinearMethod does not support weight loader v2, but
+        # transforms (specifically SharedWeightParameter) requires
+        # weight loader v2. Until UnquantizedLinearMethod supports v2, we must
+        # hack around this by getting weight loader v1 so ULM can load correctly
+        quant_method_name = self.quant_method.__class__.__name__
+        if quant_method_name not in WEIGHT_LOADER_V2_SUPPORTED:
+            if isinstance(layer, QKVCrossParallelLinear):
+                weight_loader_v1 = layer.weight_loader_v1
+            else:
+                weight_loader_v1 = layer.weight_loader
+            extra_weight_attrs["weight_loader"] = weight_loader_v1
+
+        self.quant_method.create_weights(
+            layer=layer,
+            input_size_per_partition=input_size_per_partition,
+            output_partition_sizes=output_partition_sizes,
+            input_size=input_size,
+            output_size=output_size,
+            params_dtype=params_dtype,
+            **extra_weight_attrs)
+
+        # validate schemes
+        num_partitions = len(output_partition_sizes)
+        self._validate_tfm_schemes(num_partitions)
+
+        # create submodules for weight loading
+        if len(self.input_tfms) > 0:
+            scheme_name = list(self.input_tfms.values())[0].scheme_name
+            location = list(self.input_tfms.values())[0].args.location
+            transform_name = f"{scheme_name}_{location}"
+
+            transform = HadamardTransform(self.input_tfms, layer,
+                                          weight_loader,
+                                          input_size_per_partition,
+                                          output_partition_sizes)
+            layer.register_module(transform_name, transform)
+            self.input_transform = transform
+
+        if len(self.output_tfms) > 0:
+            scheme_name = list(self.output_tfms.values())[0].scheme_name
+            location = list(self.output_tfms.values())[0].args.location
+            transform_name = f"{scheme_name}_{location}"
+
+            transform = HadamardTransform(self.output_tfms, layer,
+                                          weight_loader,
+                                          input_size_per_partition,
+                                          output_partition_sizes)
+            layer.register_module(transform_name, transform)
+            self.output_transform = transform
+
+        # compute partition ranges for slicing activations
+        starts = [0] + list(accumulate(output_partition_sizes))[:-1]
+        self.partition_ranges = list(zip(starts, output_partition_sizes))
+
+    def process_weights_after_loading(self, layer):
+        self.quant_method.process_weights_after_loading(layer)
+
+        for submodule in layer.children():
+            if isinstance(submodule, HadamardTransform):
+                submodule.process_weights_after_loading()
+
+    def apply(self,
+              layer: torch.nn.Module,
+              x: torch.Tensor,
+              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+
+        if self.input_transform is not None:
+            x = self.input_transform(x)
+
+        assert bias is None
+        x = self.quant_method.apply(layer, x, bias)
+
+        # In most cases, input transforms are preferred over output transforms
+        # (@ksayers): confirm that this is done concurrently
+        if self.output_transform is not None:
+            for part_id, (start, length) in enumerate(self.partition_ranges):
+                x[:, start:start + length] = self.output_transform(
+                    x[:, start:start + length].contiguous(), part_id=part_id)
+
+        return x
+
+    def _validate_tfm_schemes(self, num_partitions: int):
+        if len(self.input_tfms) > 0:
+            if 0 not in self.input_tfms:
+                raise ValueError("Must have same input")
+
+            for part_index in range(num_partitions):
+                if self.input_tfms[part_index] != self.input_tfms[0]:
+                    raise ValueError("Must have same input")
+
+        if len(self.output_tfms) > 0:
+            scheme_name = list(self.output_tfms.values())[0].scheme_name
+            location = list(self.output_tfms.values())[0].args.location
+
+            for tfm in self.output_tfms.values():
+                if tfm.scheme_name != scheme_name:
+                    raise ValueError("Must have same scheme name")
+                if tfm.args.location != location:
+                    raise ValueError("Must have same location")
+
+        return self.input_tfms, self.output_tfms
+
+
+def get_linear_transform_schemes(
+    layer: torch.nn.Module, layer_name: str,
+    transform_config: Optional[TransformConfig],
+    packed_modules_mapping: dict[str, list[str]]
+) -> tuple[dict[int, TransformTuple], dict[
+        int, TransformTuple]]:  # [input_transform, [output_transform, ...]]
+    # there can only be one transform input scheme per (fused) module
+    input_tfms = {}
+    output_tfms = {}
+
+    partition_names = get_layer_partition_names(layer_name,
+                                                packed_modules_mapping)
+
+    for scheme_name, scheme, args in get_schemes_args(transform_config):
+        for part_index, part_name in enumerate(partition_names):
+            if is_match(part_name, layer, args.targets,
+                        args.ignore) and args.is_online():
+                if args.location == TransformLocation.INPUT:
+                    input_tfms[part_index] = TransformTuple(
+                        scheme_name, scheme, args)
+
+                elif args.location == TransformLocation.OUTPUT:
+                    output_tfms[part_index] = TransformTuple(
+                        scheme_name, scheme, args)
+
+                else:
+                    raise ValueError(f"Cannot apply `{args.location}` "
+                                     f"transform to `{layer_name}`")
+
+    return (input_tfms, output_tfms)
+
+
+def get_schemes_args(
+    transform_config: Optional[TransformConfig]
+) -> Generator[tuple[str, TransformScheme, TransformArgs]]:
+    if transform_config is None:
+        return
+
+    for scheme_name, scheme in transform_config.config_groups.items():
+        for args in scheme.apply:
+            yield (scheme_name, scheme, args)
+
+
+def get_layer_partition_names(
+        layer_name: str, packed_modules_mapping: dict[str,
+                                                      list[str]]) -> list[str]:
+    """
+    Get all partition names associated with this layer.
+    Names are returned in order of their partition indices.
+    
+    ```python
+    mapping = {"gate_up_proj", "gate_proj", "up_proj"}
+
+    assert get_layer_partition_names(
+        "mlp.gate_up_proj", mapping) == ["gate_proj", "up_proj"]
+    assert get_layer_partition_names(
+        "mlp.down_proj", mapping) == ["down_proj"]
+    """
+    for fused_suffix, part_suffixes in packed_modules_mapping.items():
+        if layer_name.endswith(fused_suffix):
+            return [
+                layer_name.removesuffix(fused_suffix) + part_suffix
+                for part_suffix in part_suffixes
+            ]
+
+    return [layer_name]

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

@@ -0,0 +1,153 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import math
+from collections.abc import Hashable
+from typing import Callable
+
+import torch
+from compressed_tensors.transform import (TransformArgs, TransformLocation,
+                                          TransformScheme)
+from torch import Tensor
+
+import vllm._custom_ops as ops
+from vllm.distributed.parallel_state import (
+    get_tensor_model_parallel_world_size)
+from vllm.model_executor.layers.linear import LinearBase
+from vllm.model_executor.layers.quantization.compressed_tensors.transform.utils import (  # noqa: E501
+    TransformTuple)
+from vllm.model_executor.layers.utils import dispatch_unquantized_gemm
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    VocabParallelEmbedding)
+from vllm.model_executor.parameter import SharedWeightParameter
+
+
+class HadamardTransform(torch.nn.Module):
+    """
+    Class which handles weight loading, postprocessing, and application of
+    transforms. Meant to be used with `CompressedTensorsLinearTransformMethod`
+    and attention transforms method (not implemented yet)
+    """
+    transforms: dict[int, TransformTuple]  # info parsed from transforms config
+    weight: SharedWeightParameter  # container for shared tensors
+
+    scales: dict[int, float]  # hadamard scale, usually sqrt(matrix.size(0))
+
+    def __init__(self, transforms: dict[int, TransformTuple],
+                 layer: torch.nn.Module, weight_loader: Callable,
+                 input_size_per_partition: int,
+                 output_partition_sizes: list[int]):
+        super().__init__()
+        self.transforms = transforms
+        self.scales = {}
+
+        if get_tensor_model_parallel_world_size() > 1:
+            raise NotImplementedError("Online transforms with tensor "
+                                      "parallelism is not supported")
+
+        # Similar to row/col parallel params, but tensors are separate
+        # to allow for loading with shared memory
+        self.weight = SharedWeightParameter(weight_loader=weight_loader)
+
+        # create shared partition data for each partition of the original weight
+        input_size = input_size_per_partition
+        for part_index, (_scheme_name, scheme,
+                         args) in self.transforms.items():
+            output_size = output_partition_sizes[part_index]
+            weight_size = self._get_weight_size(layer, scheme, args,
+                                                input_size, output_size)
+
+            data_key = self._get_data_key(scheme, weight_size)
+            self.weight.add_partition(
+                part_index,
+                data_key,
+                size=(weight_size, weight_size),
+                dtype=scheme.precision,
+            )
+
+        # validate that shared tensors and schemes are correct
+        self._validate_input_transforms()
+
+    def process_weights_after_loading(self):
+        for part_id in self.weight.partitions:
+            data = self.weight.partitions[part_id].data
+
+            # required by torch.compile
+            self.weight.process_weights_after_loading()
+
+            # precompute scale as a runtime multiply, not division
+            # do not fold into weight in order to utilize FWHT
+            self.scales[part_id] = 1 / math.sqrt(data.size(0))
+
+            # FUTURE: avoid runtime transpose by processing weights
+            # prior to apply
+
+    def forward(self, value: Tensor, part_id: int = 0) -> Tensor:
+        if part_id not in self.weight.partitions:
+            return value
+
+        # use hadacore if possible
+        if self.transforms[part_id].scheme.type == "hadamard":
+            if self.transforms[part_id].scheme.head_dim is not None:
+                weight_size = self.transforms[part_id].scheme.head_dim
+                value = value.unflatten(-1, (-1, weight_size))
+                value = ops.hadacore_transform(value)
+                value = value.flatten(-2, -1)
+
+                return value
+
+            # sylvester transforms are symmetric, inv => transpose => original
+            return ops.hadacore_transform(value)
+
+        # fall back to dense
+        else:
+            weight = self.weight.partitions[part_id]
+            weight = weight if self.transforms[
+                part_id].args.inverse else weight.T  # linear := x(W.T)
+            scale = self.scales[part_id]
+
+            if self.transforms[part_id].scheme.head_dim is not None:
+                value = value.unflatten(-1, (-1, weight.size(0)))
+                value = dispatch_unquantized_gemm()(self, value.to(
+                    weight.dtype), weight, None).to(value.dtype) * scale
+                value = value.flatten(-2, -1)
+
+                return value
+
+            return dispatch_unquantized_gemm()(self, value.to(
+                weight.dtype), weight, None).to(value.dtype) * scale
+
+    def _get_data_key(self, scheme: TransformScheme,
+                      weight_size: int) -> Hashable:
+        return (id(scheme), weight_size)
+
+    def _get_weight_size(self, layer: torch.nn.Module, scheme: TransformScheme,
+                         args: TransformArgs, input_size: int,
+                         output_size: int) -> int:
+        if scheme.head_dim is not None:
+            return scheme.head_dim
+
+        if isinstance(layer, LinearBase):
+            if args.location == TransformLocation.INPUT:
+                return input_size
+
+            elif args.location == TransformLocation.OUTPUT:
+                return output_size
+
+        elif isinstance(layer, VocabParallelEmbedding):
+            if args.location == TransformLocation.INPUT:
+                return output_size
+
+            elif args.location == TransformLocation.OUTPUT:
+                return input_size
+
+        raise ValueError()
+
+    def _validate_input_transforms(self):
+        assert len(self.transforms) > 0
+        location = list(self.transforms.values())[0].args.location
+
+        if location == TransformLocation.INPUT:
+            first_data = self.weight.partitions[0].data
+            for partition in self.weight.partitions.values():
+                if partition.data.data_ptr() != first_data.data_ptr():
+                    raise ValueError("")

vllm/model_executor/layers/quantization/compressed_tensors/transform/schemes/linear_qutlass_nvfp4.py

@@ -0,0 +1,46 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Optional
+
+import torch
+
+from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import (  # noqa: E501
+    CompressedTensorsScheme, CompressedTensorsW4A4Fp4)
+from vllm.model_executor.layers.quantization.compressed_tensors.transform.linear import (  # noqa: E501
+    CompressedTensorsLinearTransformMethod, TransformTuple)
+
+__all__ = ["is_qutlass_fp4_scheme", "QutlassNvFP4LinearMethod"]
+
+
+def is_qutlass_fp4_scheme(quant_scheme: Optional[CompressedTensorsScheme],
+                          input_tfms: dict[int, TransformTuple]) -> bool:
+    return isinstance(
+        quant_scheme,
+        (CompressedTensorsW4A4Fp4, )) and len(input_tfms) == 1 and input_tfms[
+            0].scheme.head_dim == quant_scheme.group_size
+
+
+class QutlassNvFP4LinearMethod(CompressedTensorsLinearTransformMethod):
+
+    def create_weights(self, layer, input_size_per_partition,
+                       output_partition_sizes, input_size, output_size,
+                       params_dtype, **extra_weight_attrs):
+        # initializes fp4 qparams
+        assert isinstance(layer.scheme, (CompressedTensorsW4A4Fp4, ))
+        ret = super().create_weights(layer, input_size_per_partition,
+                                     output_partition_sizes, input_size,
+                                     output_size, params_dtype,
+                                     **extra_weight_attrs)
+
+        assert self.input_transform is not None
+        assert len(self.input_transform.weight) == 1
+        assert self.input_transform.weight[0].size(
+            0) == layer.scheme.group_size
+
+        return ret
+
+    def apply(self,
+              layer: torch.nn.Module,
+              x: torch.Tensor,
+              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+        raise NotImplementedError()

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

@@ -0,0 +1,13 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import NamedTuple
+
+from compressed_tensors.transform import TransformArgs, TransformScheme
+
+__all__ = ["TransformTuple"]
+
+
+class TransformTuple(NamedTuple):
+    scheme_name: str
+    scheme: TransformScheme
+    args: TransformArgs

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

@@ -0,0 +1,206 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from typing import Optional
+
+import torch
+
+from vllm.triton_utils import tl, triton
+
+
+def is_weak_contiguous(x: torch.Tensor):
+    strides = x.stride()
+    sizes = x.shape
+    is_not_transpose = strides[0] == 1 and (strides[1] >= max(1, sizes[0]))
+    is_transpose = strides[1] == 1 and (strides[0] >= max(1, sizes[1]))
+    return is_transpose or is_not_transpose
+
+
+@triton.jit
+def scaled_mm_kernel(a_ptr, b_ptr, scale_a_ptr, scale_b_ptr, c_ptr, bias_ptr,
+                     M, N, K, stride_am, stride_ak, stride_bk, stride_bn,
+                     stride_cm, stride_cn, ACCUMULATOR_DTYPE: tl.constexpr,
+                     BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_N: tl.constexpr,
+                     BLOCK_SIZE_K: tl.constexpr,
+                     BLOCK_SIZE_SCALE_A: tl.constexpr,
+                     BLOCK_SIZE_SCALE_B: tl.constexpr):
+    pid = tl.program_id(axis=0)
+
+    num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
+
+    pid_m = pid // num_pid_n
+    pid_n = pid % num_pid_n
+
+    accumulator_dtype = ACCUMULATOR_DTYPE
+    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N),
+                           dtype=accumulator_dtype)
+
+    # NOTE: Some tensor inputs are so large, they will cause int32 overflow
+    # so it is necessary to use tl.int64 for all the offsets, else SEGV will
+    # eventually occur.
+
+    # Offsets and masks.
+    offsets_am = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
+    masks_am = offsets_am < M
+
+    offsets_bn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)
+    masks_bn = offsets_bn < N
+
+    offsets_k = tl.arange(0, BLOCK_SIZE_K).to(tl.int64)
+    offsets_a = (stride_am * offsets_am[:, None] +
+                 stride_ak * offsets_k[None, :])
+    offsets_b = (stride_bk * offsets_k[:, None] +
+                 stride_bn * offsets_bn[None, :])
+
+    # NOTE: BLOCK_SIZE_SCALE_A could be 1 or BLOCK_SIZE_M, so need to create
+    # appropriate offsets and masks for each case. Same goes for
+    # BLOCK_SIZE_SCALE_B.
+    offsets_scale_am = (tl.arange(0, BLOCK_SIZE_SCALE_A) +
+                        (BLOCK_SIZE_SCALE_A > 1) * pid_m * BLOCK_SIZE_M)
+    masks_scale_am = offsets_scale_am < M
+
+    offsets_scale_bn = (tl.arange(0, BLOCK_SIZE_SCALE_B) +
+                        (BLOCK_SIZE_SCALE_B > 1) * pid_n * BLOCK_SIZE_N)
+    masks_scale_bn = offsets_scale_bn < N
+
+    a_ptrs = a_ptr + offsets_a
+    b_ptrs = b_ptr + offsets_b
+
+    scale_a_ptrs = scale_a_ptr + offsets_scale_am
+    scale_b_ptrs = scale_b_ptr + offsets_scale_bn
+
+    for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
+        masks_k = offsets_k < K
+        masks_a = masks_am[:, None] & masks_k[None, :]
+        a = tl.load(a_ptrs, mask=masks_a)
+
+        masks_b = masks_k[:, None] & masks_bn[None, :]
+        b = tl.load(b_ptrs, mask=masks_b)
+
+        # Accumulate results.
+        accumulator = tl.dot(a, b, accumulator, out_dtype=accumulator_dtype)
+
+        offsets_k += BLOCK_SIZE_K
+        a_ptrs += BLOCK_SIZE_K * stride_ak
+        b_ptrs += BLOCK_SIZE_K * stride_bk
+
+    # Apply scale at end.
+    masks_scale_a = masks_scale_am[:, None] & (tl.arange(0, 1) < 1)[:, None]
+    scale_a = tl.load(scale_a_ptrs[:, None], masks_scale_a)
+    # Need to broadcast to the appropriate size, if scale_a is already
+    # (BLOCK_SIZE_M, 1) then it will broadcast to its own shape. Same goes
+    # for scale_b below.
+    scale_a = scale_a.broadcast_to((BLOCK_SIZE_M, 1))
+    accumulator = scale_a * accumulator.to(tl.float32)
+
+    masks_scale_b = masks_scale_bn[:, None] & (tl.arange(0, 1) < 1)[None, :]
+    scale_b = tl.load(scale_b_ptrs[:, None], masks_scale_b)
+    scale_b = scale_b.broadcast_to((BLOCK_SIZE_N, 1))
+    accumulator = scale_b.T * accumulator.to(tl.float32)
+
+    # Convert to output format.
+    c = accumulator.to(c_ptr.type.element_ty)
+
+    # Add bias, it's already in output format, so add it after conversion.
+    if bias_ptr:
+        offsets_bias = offsets_bn
+        bias_ptrs = bias_ptr + offsets_bias
+        bias_mask = offsets_bias < N
+        bias = tl.load(bias_ptrs, bias_mask)
+        c += bias
+
+    # Save output
+    offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
+    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)
+    offs_cm = offs_cm.to(tl.int64)
+    offs_cn = offs_cn.to(tl.int64)
+    c_ptrs = (c_ptr + stride_cm * offs_cm[:, None] +
+              stride_cn * offs_cn[None, :])
+    c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
+
+    tl.store(c_ptrs, c, mask=c_mask)
+
+
+# input   - [M, K]
+# weight - [K, N]
+def triton_scaled_mm(input: torch.Tensor,
+                     weight: torch.Tensor,
+                     scale_a: torch.Tensor,
+                     scale_b: torch.Tensor,
+                     out_dtype: type[torch.dtype],
+                     bias: Optional[torch.Tensor] = None,
+                     block_size_m: int = 32,
+                     block_size_n: int = 32,
+                     block_size_k: int = 32,
+                     use_heuristic=True) -> torch.Tensor:
+    M, K = input.shape
+    N = weight.shape[1]
+
+    assert N > 0 and K > 0 and M > 0
+    assert weight.shape[0] == K
+    assert input.dtype == weight.dtype
+
+    scale_a = scale_a.reshape(-1, 1) if scale_a.dim() <= 1 else scale_a
+    scale_b = scale_b.reshape(-1, 1) if scale_b.dim() <= 1 else scale_b
+
+    assert scale_a.dtype == scale_b.dtype and scale_a.is_floating_point()
+    assert scale_a.shape[1] == 1 and (scale_a.shape[0] == 1
+                                      or scale_a.shape[0] == M)
+    assert scale_b.shape[1] == 1 and (scale_b.shape[0] == 1
+                                      or scale_b.shape[0] == N)
+    assert out_dtype.is_floating_point
+    assert bias is None or bias.is_floating_point()
+    assert is_weak_contiguous(input)
+    assert is_weak_contiguous(weight)
+
+    grid = lambda META: (triton.cdiv(M, META['BLOCK_SIZE_M']) * triton.cdiv(
+        N, META['BLOCK_SIZE_N']), )
+
+    result = torch.empty((M, N), dtype=out_dtype, device=input.device)
+
+    has_scalar = lambda x: x.shape[0] == 1 and x.shape[1] == 1
+
+    if use_heuristic:
+        is_small_N = N < 8192
+        next_power_of_2_M = max(32, triton.next_power_of_2(M))
+        if next_power_of_2_M <= 32:
+            tile_shape = (64, 64, 256) if is_small_N else (64, 128, 256)
+        elif next_power_of_2_M <= 64:
+            tile_shape = (64, 64, 256)
+        elif next_power_of_2_M <= 128:
+            tile_shape = (64, 128, 128)
+        else:
+            tile_shape = (128, 128, 128)
+
+    block_size_m, block_size_n, block_size_k = tile_shape
+
+    block_size_sa = 1 if has_scalar(scale_a) else block_size_m
+    block_size_sb = 1 if has_scalar(scale_b) else block_size_n
+
+    accumulator_dtype = tl.float32 if input.is_floating_point() else tl.int32
+
+    # A = input, B = weight, C = result
+    # A = M x K, B = K x N, C = M x N
+    scaled_mm_kernel[grid](input,
+                           weight,
+                           scale_a,
+                           scale_b,
+                           result,
+                           bias,
+                           M,
+                           N,
+                           K,
+                           input.stride(0),
+                           input.stride(1),
+                           weight.stride(0),
+                           weight.stride(1),
+                           result.stride(0),
+                           result.stride(1),
+                           accumulator_dtype,
+                           BLOCK_SIZE_M=block_size_m,
+                           BLOCK_SIZE_N=block_size_n,
+                           BLOCK_SIZE_K=block_size_k,
+                           BLOCK_SIZE_SCALE_A=block_size_sa,
+                           BLOCK_SIZE_SCALE_B=block_size_sb)
+
+    return result.to(out_dtype)