Initial import: grid-bot — grid trading bot for BTC-USDT on Cifra Markets

kronos-venv/lib/python3.12/site-packages/torch/ao/nn/sparse/__init__.py

@@ -0,0 +1 @@
+from . import quantized

kronos-venv/lib/python3.12/site-packages/torch/ao/nn/sparse/quantized/__init__.py

@@ -0,0 +1,10 @@
+from torch.ao.nn.sparse.quantized import dynamic
+
+from .linear import Linear, LinearPackedParams
+
+
+__all__ = [
+    "dynamic",
+    "Linear",
+    "LinearPackedParams",
+]

kronos-venv/lib/python3.12/site-packages/torch/ao/nn/sparse/quantized/dynamic/__init__.py

@@ -0,0 +1,6 @@
+from .linear import Linear
+
+
+__all__ = [
+    "Linear",
+]

kronos-venv/lib/python3.12/site-packages/torch/ao/nn/sparse/quantized/dynamic/linear.py

@@ -0,0 +1,203 @@
+# mypy: allow-untyped-defs
+
+import torch
+import torch.ao.nn.intrinsic as nni
+from torch.ao.nn.quantized.modules.utils import (
+    _hide_packed_params_repr,
+    _quantize_weight,
+)
+from torch.ao.nn.sparse.quantized import linear
+from torch.ao.nn.sparse.quantized.utils import LinearBlockSparsePattern
+
+
+__all__ = ["Linear"]
+
+
+class Linear(torch.nn.Module):
+    r"""
+    A dynamically quantized sparse linear module with float tensor as inputs and outputs.
+    """
+
+    _version = 1
+    _op_type = "sparse_dynamic"
+    _FLOAT_MODULE = torch.nn.Linear
+
+    def __init__(
+        self,
+        in_features,
+        out_features,
+        row_block_size,
+        col_block_size,
+        bias=True,
+        dtype=torch.qint8,
+    ):
+        super().__init__()
+
+        if dtype != torch.qint8:
+            raise NotImplementedError(
+                "Only QINT8 is supported for Sparse Quantized Linear Dynamic"
+            )
+
+        self.in_features = in_features
+        self.out_features = out_features
+
+        if bias:
+            bias = torch.zeros(self.out_features, dtype=torch.float)
+        else:
+            bias = None
+
+        qweight = torch._empty_affine_quantized(
+            [out_features, in_features], scale=1, zero_point=0, dtype=torch.qint8
+        )
+        self._packed_params = linear.LinearPackedParams(
+            row_block_size=row_block_size, col_block_size=col_block_size, dtype=dtype
+        )
+        self._packed_params.set_weight_bias(
+            qweight, bias, row_block_size, col_block_size
+        )
+
+    def _get_name(self):
+        return "SparseQuantizedDynamicLinear"
+
+    def extra_repr(self):
+        return f"in_features={self.in_features}, out_features={self.out_features}, qscheme={self.weight().qscheme()}"
+
+    def __repr__(self):
+        return _hide_packed_params_repr(self, linear.LinearPackedParams)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return torch.ops.sparse.qlinear_dynamic(x, self._packed_params._packed_params)
+
+    def _save_to_state_dict(self, destination, prefix, keep_vars):
+        super()._save_to_state_dict(destination, prefix, keep_vars)
+        destination[prefix + "op_type"] = self._op_type
+
+    def _load_from_state_dict(
+        self,
+        state_dict,
+        prefix,
+        local_metadata,
+        strict,
+        missing_keys,
+        unexpected_keys,
+        error_msgs,
+    ):
+        op_type = int(state_dict[prefix + "op_type"])
+        if op_type != "sparse":
+            raise AssertionError(
+                f"Cannot load from op_type [{op_type}], expecting [{self._op_type}]"
+            )
+        state_dict.pop(prefix + "op_type")
+
+        version = local_metadata.get("version", None)
+        if version is not None and version > self._version:
+            raise AssertionError(f"version {version} > self._version {self._version}")
+
+        # Is this code valid? In old quantization it seemed to be used to load
+        # older model
+        weight = state_dict.pop(prefix + "weight")
+        bias = state_dict.pop(prefix + "bias")
+        state_dict.update(
+            {
+                prefix + "_packed_params.weight": weight,
+                prefix + "_packed_params.bias": bias,
+            }
+        )
+
+        super()._load_from_state_dict(
+            state_dict,
+            prefix,
+            local_metadata,
+            False,
+            missing_keys,
+            unexpected_keys,
+            error_msgs,
+        )
+
+    def _weight_bias(self):
+        return self._packed_params._weight_bias()
+
+    def weight(self):
+        return self._weight_bias()[0]
+
+    def bias(self):
+        return self._weight_bias()[1]
+
+    def set_weight_bias(
+        self,
+        w: torch.Tensor,
+        b: torch.Tensor | None,
+        row_block_size: int | None,
+        col_block_size: int | None,
+    ) -> None:
+        if row_block_size is None or col_block_size is None:
+            raise AssertionError(
+                f"row_block_size and col_block_size must not be None, got {row_block_size=}, {col_block_size=}"
+            )
+        self.out_features = w.shape[0]
+        self.in_features = w.shape[1]
+        self._packed_params.set_weight_bias(w, b, row_block_size, col_block_size)
+
+    @classmethod
+    def from_float(cls, mod, use_precomputed_fake_quant=False):
+        r"""Create a quantized sparse dynamic module from a float module.
+
+        We only care about the convert at this stage, no need for observers just yet.
+        """
+        if type(mod) is not cls._FLOAT_MODULE:
+            raise AssertionError(
+                " nnq."
+                + cls.__name__
+                + ".from_float only works for "
+                + cls._FLOAT_MODULE.__name__
+            )
+        # TODO: Need to add options to qconfig to avoid the calibration.
+        # TODO: Add calibration for the sparsity
+        if not hasattr(mod, "qconfig"):
+            raise AssertionError("Input float module must have qconfig defined")
+        if type(mod) is nni.LinearReLU:
+            mod = mod[0]
+        # pyrefly: ignore [missing-attribute]
+        if mod.qconfig is not None and mod.qconfig.weight is not None:
+            # pyrefly: ignore [not-callable]
+            weight_observer = mod.qconfig.weight()
+        else:
+            # We have the circular import issues if we import the qconfig in the beginning of this file:
+            # https://github.com/pytorch/pytorch/pull/24231. The current workaround is to postpone the
+            # import until we need it.
+            from torch.ao.quantization.qconfig import default_dynamic_qconfig
+
+            weight_observer = default_dynamic_qconfig.weight()
+
+        # It is important to multiply by the mask BEFORE calling the `weight_observer`
+        # TODO (zaf): Mask might not be part of the qconfig (T83295194)
+        weight = mod.weight
+        if getattr(mod.qconfig, "mask", False):
+            weight = mod.qconfig.mask * mod.weight
+
+        weight_observer(weight)
+        dtype = weight_observer.dtype
+        if dtype != torch.qint8:
+            raise AssertionError(
+                f"Weight observer must have dtype torch.qint8, got {dtype}"
+            )
+        _w_sc, w_zp = weight_observer.calculate_qparams()
+        if isinstance(w_zp, torch.Tensor):
+            if torch.any(w_zp.bool()):
+                raise AssertionError("All weight zero points must map to 0")
+        else:
+            if w_zp != 0:
+                raise AssertionError(f"Weight zero point must map to 0, got {w_zp}")
+        qweight = _quantize_weight(weight.float(), weight_observer)
+
+        row_block_size, col_block_size = LinearBlockSparsePattern.block_size()
+        qlinear = cls(
+            mod.in_features,
+            mod.out_features,
+            row_block_size,
+            col_block_size,
+            dtype=dtype,
+        )
+        # pyrefly: ignore [bad-argument-type]
+        qlinear.set_weight_bias(qweight, mod.bias, row_block_size, col_block_size)
+        return qlinear

kronos-venv/lib/python3.12/site-packages/torch/ao/nn/sparse/quantized/linear.py

@@ -0,0 +1,300 @@
+# mypy: allow-untyped-defs
+
+import torch
+from torch.ao.nn.quantized.modules.utils import (
+    _hide_packed_params_repr,
+    _quantize_weight,
+)
+
+
+__all__ = ["LinearPackedParams", "Linear"]
+
+
+# TODO (zaf): Inherit from `quantized.LinearPackedParams` (T83294430)
+class LinearPackedParams(torch.nn.Module):
+    _version = 1
+
+    def __init__(self, row_block_size=1, col_block_size=4, dtype=torch.qint8):
+        super().__init__()
+
+        if dtype != torch.qint8:
+            raise NotImplementedError("Linear prepacking only supports QINT8")
+        self.dtype = dtype
+        wq = torch._empty_affine_quantized(
+            [1, 1], scale=1.0, zero_point=0, dtype=torch.qint8
+        )
+        self.set_weight_bias(wq, None, row_block_size, col_block_size)
+
+    def _get_name(self):
+        return "SparseQuantizedLinearPackedParams"
+
+    @torch.jit.export
+    def set_weight_bias(
+        self,
+        weight: torch.Tensor,
+        bias: torch.Tensor | None,
+        row_block_size: int | None,
+        col_block_size: int | None,
+    ) -> None:
+        if row_block_size is None or col_block_size is None:
+            raise AssertionError(
+                "row_block_size and col_block_size must not be None, got "
+                f"row_block_size={row_block_size}, col_block_size={col_block_size}"
+            )
+        self._packed_params = torch.ops.sparse.qlinear_prepack(
+            weight, bias, row_block_size, col_block_size
+        )
+
+    @torch.jit.export
+    def _weight_bias(self):
+        (weight, bias, block_sizes) = torch.ops.sparse.qlinear_unpack(
+            self._packed_params
+        )
+        return (weight, bias, block_sizes[0], block_sizes[1])
+
+    def forward(self, x):
+        return x
+
+    def _save_to_state_dict(self, destination, prefix, keep_vars):
+        super()._save_to_state_dict(destination, prefix, keep_vars)
+        destination[prefix + "dtype"] = self.dtype
+        destination[prefix + "_packed_params"] = self._weight_bias()
+
+    def _load_from_state_dict(
+        self,
+        state_dict,
+        prefix,
+        local_metadata,
+        strict,
+        missing_keys,
+        unexpected_keys,
+        error_msgs,
+    ):
+        version = local_metadata.get("version", None)
+        if version is not None and version > self._version:
+            raise AssertionError(f"version {version} > self._version {self._version}")
+
+        self.dtype = state_dict.pop(prefix + "dtype")
+        weight, bias, row_block_size, col_block_size = state_dict.pop(
+            prefix + "_packed_params"
+        )
+        self.set_weight_bias(weight, bias, row_block_size, col_block_size)
+
+        super()._load_from_state_dict(
+            state_dict,
+            prefix,
+            local_metadata,
+            False,
+            missing_keys,
+            unexpected_keys,
+            error_msgs,
+        )
+
+    @torch.jit.export
+    def __getstate__(self):
+        return self._packed_params, self.training, self.dtype
+
+    @torch.jit.export
+    def __setstate__(self, state):
+        (self._packed_params, self.training, self.dtype) = state
+
+    def __repr__(self):
+        return self._weight_bias().__repr__()
+
+
+# TODO (zaf): Inherit from `quantized.Linear` (T83294430)
+class Linear(torch.nn.Module):
+    r"""
+    A quantized sparse linear module with quantized tensor as inputs and outputs.
+    """
+
+    _version = 1
+    _FLOAT_MODULE = torch.nn.Linear
+
+    def __init__(
+        self,
+        in_features,
+        out_features,
+        row_block_size,
+        col_block_size,
+        bias=True,
+        dtype=torch.qint8,
+    ):
+        super().__init__()
+
+        if dtype != torch.qint8:
+            raise NotImplementedError(
+                "Only QINT8 is supported for Sparse Quantized Linear"
+            )
+
+        self.in_features = in_features
+        self.out_features = out_features
+
+        if bias:
+            bias = torch.zeros(self.out_features, dtype=torch.float)
+        else:
+            bias = None
+
+        qweight = torch._empty_affine_quantized(
+            [out_features, in_features], scale=1, zero_point=0, dtype=torch.qint8
+        )
+        self._packed_params = LinearPackedParams(
+            row_block_size=row_block_size, col_block_size=col_block_size, dtype=dtype
+        )
+        self._packed_params.set_weight_bias(
+            qweight, bias, row_block_size, col_block_size
+        )
+        self.scale = 1.0
+        self.zero_point = 0
+
+    @classmethod
+    def _get_name(cls):
+        return "SparseQuantizedLinear"
+
+    def extra_repr(self):
+        return (
+            f"in_features={self.in_features}, out_features={self.out_features}, scale={self.scale}, "
+            f"zero_point={self.zero_point}, qscheme={self.weight().qscheme()}"
+        )
+
+    def __repr__(self):
+        return _hide_packed_params_repr(self, LinearPackedParams)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return torch.ops.sparse.qlinear(
+            x, self._packed_params._packed_params, self.scale, self.zero_point
+        )
+
+    def _save_to_state_dict(self, destination, prefix, keep_vars):
+        super()._save_to_state_dict(destination, prefix, keep_vars)
+        destination[prefix + "scale"] = torch.tensor(self.scale)
+        destination[prefix + "zero_point"] = torch.tensor(self.zero_point)
+
+    def _load_from_state_dict(
+        self,
+        state_dict,
+        prefix,
+        local_metadata,
+        strict,
+        missing_keys,
+        unexpected_keys,
+        error_msgs,
+    ):
+        self.scale = float(state_dict[prefix + "scale"])
+        state_dict.pop(prefix + "scale")
+
+        self.zero_point = int(state_dict[prefix + "zero_point"])
+        state_dict.pop(prefix + "zero_point")
+
+        state_dict.pop(prefix + "op_type")
+
+        version = local_metadata.get("version", None)
+        if version is not None and version > self._version:
+            raise AssertionError(f"version {version} > self._version {self._version}")
+
+        super()._load_from_state_dict(
+            state_dict,
+            prefix,
+            local_metadata,
+            False,
+            missing_keys,
+            unexpected_keys,
+            error_msgs,
+        )
+
+    def _weight_bias(self):
+        return self._packed_params._weight_bias()
+
+    def weight(self):
+        return self._weight_bias()[0]
+
+    def bias(self):
+        return self._weight_bias()[1]
+
+    def set_weight_bias(
+        self,
+        w: torch.Tensor,
+        b: torch.Tensor | None,
+        row_block_size: int | None,
+        col_block_size: int | None,
+    ) -> None:
+        if row_block_size is None or col_block_size is None:
+            raise AssertionError(
+                "row_block_size and col_block_size must not be None, "
+                f"got row_block_size={row_block_size}, col_block_size={col_block_size}"
+            )
+        self._packed_params.set_weight_bias(w, b, row_block_size, col_block_size)
+
+    @classmethod
+    def from_float(cls, mod, use_precomputed_fake_quant=False):
+        r"""Create a quantized sparse module from a float module.
+
+        We only care about the convert at this stage, no need for observers just yet.
+
+        TODO(zaf): Need to add the sparse params to the qconfig
+        """
+        if type(mod) is not cls._FLOAT_MODULE:
+            raise AssertionError(
+                cls._get_name()
+                + ".from_float only works for "
+                + cls._FLOAT_MODULE.__name__
+            )
+        if not hasattr(mod, "sparse_params"):
+            raise AssertionError(
+                "Expecting the Linear to have `sparse_params`. Make sure you have provided arguments "
+                'in the `sparsifier.squash_mask(params_to_save=("sparse_block_shape",))` method.'
+            )
+        sparse_block_shape = mod.sparse_params.get("sparse_block_shape", None)  # type: ignore[operator, union-attr]
+        if not isinstance(sparse_block_shape, (tuple, list)):
+            raise AssertionError(
+                f"sparse_block_shape must be tuple or list, got {type(sparse_block_shape)}"
+            )
+        if len(sparse_block_shape) != 2:
+            raise AssertionError(
+                f"sparse_block_shape must have length 2, got {len(sparse_block_shape)}"
+            )
+        # TODO: Need to add options to qconfig to avoid the calibration.
+        # TODO: Add calibration for the sparsity
+        if not hasattr(mod, "qconfig"):
+            raise AssertionError("Input float module must have qconfig defined")
+        activation_post_process = mod.activation_post_process
+        weight_post_process = mod.qconfig.weight()  # type: ignore[operator, union-attr]
+
+        # Assumption is that the weight is already sparsified by the
+        # `sparsifier.convert`
+        weight = mod.weight
+
+        weight_post_process(weight)
+        dtype = weight_post_process.dtype
+        act_scale, act_zp = activation_post_process.calculate_qparams()  # type: ignore[operator, union-attr]
+        if dtype != torch.qint8:
+            raise AssertionError(
+                f"Weight observer must have dtype torch.qint8, got {dtype}"
+            )
+        w_sc, w_zp = weight_post_process.calculate_qparams()
+        if isinstance(w_zp, torch.Tensor):
+            if torch.any(w_zp.bool()):
+                raise AssertionError("All weight zero points must map to 0")
+        else:
+            if w_zp != 0:
+                raise AssertionError(f"Weight zero point must map to 0, got {w_zp}")
+        qweight = _quantize_weight(weight.float(), weight_post_process)
+
+        row_block_size = mod.sparse_params["sparse_block_shape"][0]  # type: ignore[index]
+        col_block_size = mod.sparse_params["sparse_block_shape"][1]  # type: ignore[index]
+        qlinear = cls(
+            mod.in_features,
+            mod.out_features,
+            row_block_size,
+            col_block_size,
+            dtype=dtype,
+        )
+        qlinear.set_weight_bias(
+            qweight,
+            mod.bias,
+            row_block_size,  # type: ignore[arg-type]
+            col_block_size,  # type: ignore[arg-type]
+        )
+        qlinear.scale = float(act_scale)
+        qlinear.zero_point = int(act_zp)
+        return qlinear

kronos-venv/lib/python3.12/site-packages/torch/ao/nn/sparse/quantized/utils.py

@@ -0,0 +1,66 @@
+import threading
+
+
+__all__ = ["LinearBlockSparsePattern"]
+
+
+def _is_valid_linear_block_sparse_pattern(
+    row_block_size: int, col_block_size: int
+) -> bool:
+    return (row_block_size == 1 and col_block_size == 4) or (
+        row_block_size == 8 and col_block_size == 1
+    )
+
+
+# This is a stop-gap measure as current flow does not allow module
+# specific block sparse pattern.
+# In fact there is no way to convey sparse pattern via module config
+# of quantization flow. Thus using the global context to convey
+# sparsity pattern.
+# Once the flow supports it, this should be removed.
+class LinearBlockSparsePattern:
+    rlock = threading.RLock()
+    row_block_size: int = 1
+    col_block_size: int = 4
+    prev_row_block_size: int = 1
+    prev_col_block_size: int = 4
+
+    def __init__(self, row_block_size: int = 1, col_block_size: int = 4):
+        if not _is_valid_linear_block_sparse_pattern(row_block_size, col_block_size):
+            raise AssertionError(
+                f"Invalid linear block sparse pattern: "
+                f"row_block_size={row_block_size}, col_block_size={col_block_size}"
+            )
+        LinearBlockSparsePattern.rlock.acquire()
+        LinearBlockSparsePattern.prev_row_block_size = (
+            LinearBlockSparsePattern.row_block_size
+        )
+        LinearBlockSparsePattern.prev_col_block_size = (
+            LinearBlockSparsePattern.col_block_size
+        )
+        LinearBlockSparsePattern.row_block_size = row_block_size
+        LinearBlockSparsePattern.col_block_size = col_block_size
+
+    def __enter__(self) -> None:
+        pass
+
+    def __exit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_value: BaseException | None,
+        backtrace: object | None,
+    ) -> None:
+        LinearBlockSparsePattern.row_block_size = (
+            LinearBlockSparsePattern.prev_row_block_size
+        )
+        LinearBlockSparsePattern.col_block_size = (
+            LinearBlockSparsePattern.prev_col_block_size
+        )
+        LinearBlockSparsePattern.rlock.release()
+
+    @staticmethod
+    def block_size() -> tuple[int, int]:
+        return (
+            LinearBlockSparsePattern.row_block_size,
+            LinearBlockSparsePattern.col_block_size,
+        )