paddlex 3.0.0rc0__py3-none-any.whl → 3.0.0rc1__py3-none-any.whl

paddlex/inference/models/common/vlm/bert_padding.py

@@ -0,0 +1,127 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import operator
+from functools import reduce
+
+import paddle
+import paddle.nn.functional as F
+
+
+class IndexFirstAxis(paddle.autograd.PyLayer):
+
+    @staticmethod
+    def forward(ctx, input, indices):
+        from einops import rearrange, repeat
+
+        ctx.save_for_backward(indices)
+        assert input.ndim >= 2
+        ctx.first_axis_dim, other_shape = input.shape[0], input.shape[1:]
+        second_dim = reduce(operator.mul, other_shape, 1)
+        return paddle.take_along_axis(
+            arr=rearrange(input, "b ... -> b (...)"),
+            axis=0,
+            indices=repeat(indices, "z -> z d", d=second_dim),
+        ).reshape([-1, *other_shape])
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        """Class Attribute: torch.autograd.function.FunctionCtx.saved_tensors, can not convert, please check whether it is torch.Tensor.*/torch.autograd.function.FunctionCtx.*/torch.distributions.Distribution.* and convert manually"""
+        from einops import rearrange, repeat
+
+        (indices,) = ctx.saved_tensor()
+        assert grad_output.ndim >= 2
+        other_shape = grad_output.shape[1:]
+        grad_output = rearrange(grad_output, "b ... -> b (...)")
+        grad_input = paddle.zeros(
+            shape=[ctx.first_axis_dim, tuple(grad_output.shape)[1]],
+            dtype=grad_output.dtype,
+        )
+
+        grad_input.put_along_axis_(
+            axis=0,
+            indices=repeat(indices, "z -> z d", d=tuple(grad_output.shape)[1]),
+            values=grad_output,
+        )
+        return grad_input.reshape([ctx.first_axis_dim, *other_shape]), None
+
+
+index_first_axis = IndexFirstAxis.apply
+
+
+class IndexPutFirstAxis(paddle.autograd.PyLayer):
+
+    @staticmethod
+    def forward(ctx, values, indices, first_axis_dim):
+        ctx.save_for_backward(indices)
+        assert indices.ndim == 1
+        assert values.ndim >= 2
+        output = paddle.zeros(
+            shape=[first_axis_dim, *tuple(values.shape)[1:]], dtype=values.dtype
+        )
+        output[indices] = values
+        return output
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        """Class Attribute: torch.autograd.function.FunctionCtx.saved_tensors, can not convert, please check whether it is torch.Tensor.*/torch.autograd.function.FunctionCtx.*/torch.distributions.Distribution.* and convert manually"""
+        (indices,) = ctx.saved_tensor()
+        grad_values = grad_output[indices]
+        return grad_values, None
+
+
+index_put_first_axis = IndexPutFirstAxis.apply
+
+
+def unpad_input(hidden_states, attention_mask):
+    """
+    Arguments:
+        hidden_states: (batch, seqlen, ...)
+        attention_mask: (batch, seqlen), bool / int, 1 means valid and 0 means not valid.
+    Return:
+        hidden_states: (total_nnz, ...), where total_nnz = number of tokens in selected in attention_mask.
+        indices: (total_nnz), the indices of non-masked tokens from the flattened input sequence.
+        cu_seqlens: (batch + 1), the cumulative sequence lengths, used to index into hidden_states.
+        max_seqlen_in_batch: int
+    """
+    from einops import rearrange
+
+    seqlens_in_batch = paddle.sum(attention_mask, axis=-1, dtype="int32")
+    indices = paddle.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = paddle.max(seqlens_in_batch).item()
+    cu_seqlens = F.pad(paddle.cumsum(seqlens_in_batch, axis=0), [1, 0])
+
+    return (
+        index_first_axis(rearrange(hidden_states, "b s ... -> (b s) ..."), indices),
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+def pad_input(hidden_states, indices, batch, seqlen):
+    """
+    Arguments:
+        hidden_states: (total_nnz, ...), where total_nnz = number of tokens in selected in attention_mask.
+        indices: (total_nnz), the indices that represent the non-masked tokens of the original padded input sequence.
+        batch: int, batch size for the padded sequence.
+        seqlen: int, maximum sequence length for the padded sequence.
+    Return:
+        hidden_states: (batch, seqlen, ...)
+    """
+    from einops import rearrange
+
+    output = index_put_first_axis(hidden_states, indices, batch * seqlen)
+    return rearrange(output, "(b s) ... -> b s ...", b=batch)

paddlex/inference/models/common/vlm/distributed.py

@@ -0,0 +1,229 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import re
+from typing import Any, Union
+
+import numpy as np
+import paddle
+import paddle.distributed as distributed
+
+from .utils import device_guard
+
+world_size = distributed.get_world_size()
+
+
+def convert_file_size_to_int(size: Union[int, str]):
+    """
+    Converts a size expressed as a string with digits an unit (like `"5MB"`) to an integer (in bytes).
+    Args:
+        size (`int` or `str`): The size to convert. Will be directly returned if an `int`.
+    """
+    if isinstance(size, int):
+        return size
+    if size.upper().endswith("GIB"):
+        return int(size[:-3]) * (2**30)
+    if size.upper().endswith("MIB"):
+        return int(size[:-3]) * (2**20)
+    if size.upper().endswith("KIB"):
+        return int(size[:-3]) * (2**10)
+    if size.upper().endswith("GB"):
+        int_size = int(size[:-2]) * (10**9)
+        return int_size // 8 if size.endswith("b") else int_size
+    if size.upper().endswith("MB"):
+        int_size = int(size[:-2]) * (10**6)
+        return int_size // 8 if size.endswith("b") else int_size
+    if size.upper().endswith("KB"):
+        int_size = int(size[:-2]) * (10**3)
+        return int_size // 8 if size.endswith("b") else int_size
+    raise ValueError(
+        "`size` is not in a valid format. Use an integer followed by the unit, e.g., '5GB'."
+    )
+
+
+def reduce_tensor(tensor, buffer_size="32MiB"):
+    if tensor.dtype == paddle.int8:
+        numel = np.prod(tensor.shape)
+    else:
+        numel = int(paddle.numel(tensor).item())
+    buffer_size = convert_file_size_to_int(buffer_size)
+    tensor.reshape_([-1])
+
+    send_size = buffer_size // dtype_byte_size(tensor.dtype)
+
+    for x in range(0, numel, send_size):
+        part_tensor = tensor[x : min(numel, x + send_size)]
+        yield part_tensor, (x, min(numel, x + send_size))
+
+
+def dtype_byte_size(dtype):
+    """
+    Returns the size (in bytes) occupied by one parameter of type `dtype`.
+    """
+    if dtype == paddle.bool:
+        return 1 / 8
+    bit_search = re.search(r"[^\d](\d+)$", str(dtype))
+    if bit_search is None:
+        raise ValueError(f"`dtype` is not a valid dtype: {dtype}.")
+    bit_size = int(bit_search.groups()[0])
+    return bit_size // 8
+
+
+@paddle.no_grad()
+def distributed_gather(tensor: Any, dst: int = 0, group=None, offload=False) -> Any:
+    try:
+        if isinstance(tensor, (tuple, list)):
+            return type(tensor)(
+                distributed_gather(t, dst, group, offload) for t in tensor
+            )
+        if isinstance(tensor, dict):
+            return {
+                k: distributed_gather(v, dst, group, offload) for k, v in tensor.items()
+            }
+
+        output_tensors = None
+
+        is_dst = dst == distributed.get_rank(group=group)
+        if is_dst:
+            if offload:
+                output_tensors = [
+                    [] for _ in range(distributed.get_world_size(group=group))
+                ]
+            else:
+                output_tensors = [
+                    paddle.empty_like(tensor)
+                    for _ in range(distributed.get_world_size(group=group))
+                ]
+                output_tensors = [
+                    t if len(t.shape) > 0 else t[None] for t in output_tensors
+                ]
+
+        if offload:
+            origin_shape = tensor.shape
+            tensor.reshape_([-1])
+
+            for slice_tensor, index in reduce_tensor(tensor):
+                slice_output_tensors = None
+                if distributed.get_rank(group=group) == dst:
+                    slice_output_tensors = [
+                        paddle.empty_like(slice_tensor)
+                        for _ in range(distributed.get_world_size(group=group))
+                    ]
+                paddle.distributed.communication.stream.gather(
+                    slice_tensor,
+                    slice_output_tensors,
+                    dst=group.ranks[dst] if group else dst,
+                    group=group,
+                    sync_op=True,
+                    use_calc_stream=False,
+                )
+
+                if is_dst:
+                    for i in range(len(output_tensors)):
+                        output_tensors[i].append(slice_output_tensors[i].cpu().numpy())
+
+            tensor.reshape_(origin_shape)
+            if is_dst:
+                with device_guard("cpu"):
+                    new_output_tensors = []
+                    for x in output_tensors:
+                        t = np.concatenate(x)
+                        t = t.reshape(origin_shape)
+                        new_output_tensors.append(t)
+                    output_tensors = new_output_tensors
+
+        else:
+            paddle.distributed.communication.stream.gather(
+                tensor,
+                output_tensors,
+                dst=group.ranks[dst] if group else dst,
+                group=group,
+                sync_op=True,
+                use_calc_stream=False,
+            )
+
+        return output_tensors
+
+    except AssertionError:
+        raise AssertionError("Not currently using distributed training")
+
+
+@paddle.no_grad()
+def distributed_allgather(tensor: Any, group=None, offload=False):
+    """nested all gather function with offload
+
+    Args:
+        tensor (Any): the desired tensor, list of tensor, dict of tensor to allgather.
+        group (_type_, optional): the communication group. Defaults to None.
+        offload (bool, optional): If True, we offload the received tensor to cpu/(numpy). Defaults to False.
+
+    Raises:
+        AssertionError: Unexpected errors.
+
+    Returns:
+        tensor list: list of all gathered tensors
+    """
+    try:
+        if isinstance(tensor, (tuple, list)):
+            return type(tensor)(
+                distributed_allgather(t, group, offload) for t in tensor
+            )
+        if isinstance(tensor, dict):
+            return {
+                k: distributed_allgather(v, group, offload) for k, v in tensor.items()
+            }
+
+        output_tensors = []
+
+        if offload:
+            with device_guard("cpu"):
+                output_tensors = [
+                    paddle.empty_like(tensor)
+                    for _ in range(distributed.get_world_size(group))
+                ]
+        else:
+            output_tensors = [
+                paddle.empty_like(tensor)
+                for _ in range(distributed.get_world_size(group))
+            ]
+
+        output_tensors = [t if len(t.shape) > 0 else t[None] for t in output_tensors]
+
+        if offload:
+            origin_shape = tensor.shape
+            tensor.reshape_([-1])
+            for x in output_tensors:
+                x.reshape_([-1])
+
+            for slice_tensor, index in reduce_tensor(tensor):
+                slice_output_tensors = [
+                    paddle.empty_like(slice_tensor)
+                    for _ in range(distributed.get_world_size(group))
+                ]
+                distributed.all_gather(slice_output_tensors, slice_tensor, group=group)
+                for x, y in zip(slice_output_tensors, output_tensors):
+                    with device_guard("cpu"):
+                        y[index[0] : index[1]] = x.cpu()
+
+            tensor.reshape_(origin_shape)
+            for x in output_tensors:
+                x.reshape_(origin_shape)
+
+        else:
+            distributed.all_gather(output_tensors, tensor)
+
+        return output_tensors
+
+    except AssertionError:
+        raise AssertionError("Not currently using distributed training")

paddlex/inference/models/common/vlm/flash_attn_utils.py

@@ -0,0 +1,119 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import paddle
+
+
+def is_flash_attn_available():
+    try:
+        import os
+
+        if "npu" in paddle.get_device():  # NOTE: flash attn has not been tested yet
+            for lib in os.listdir(os.getenv("CUSTOM_DEVICE_ROOT")):
+                if lib.endswith(".so"):
+                    paddle.utils.cpp_extension.extension_utils.load_op_meta_info_and_register_op(
+                        lib
+                    )
+            from paddle.base import core
+
+            def flash_attention_npu(
+                query,
+                key,
+                value,
+                dropout=0.0,
+                causal=False,
+                return_softmax=False,
+                *,
+                fixed_seed_offset=None,
+                rng_name="",
+                training=True,
+                name=None,
+                attn_mask=None,
+                is_varlen=False,
+                batch_size=None,
+                seq_length=None,
+            ):
+
+                is_triangle_upper_mask = True
+
+                if is_varlen:
+                    if len(query.shape) == 4:
+                        B, S, N, D = query.shape
+                        query = query.reshape([B * S, N, D])
+                        key = key.reshape([B * S, N, D])
+                        value = value.reshape([B * S, N, D])
+                    else:
+                        assert batch_size is not None
+                        assert seq_length is not None
+                        B = batch_size
+                        S = seq_length
+                    actual_seq_q_len = actual_seq_kv_len = list(range(S, B * S + S, S))
+                else:
+                    actual_seq_q_len = actual_seq_kv_len = []
+
+                out = core.eager._run_custom_op(
+                    "flash_attention_npu",
+                    query,
+                    key,
+                    value,
+                    fixed_seed_offset,
+                    attn_mask,
+                    actual_seq_q_len,
+                    actual_seq_kv_len,
+                    dropout,
+                    causal,
+                    return_softmax,
+                    not training,
+                    is_triangle_upper_mask,
+                    is_varlen,
+                )[0]
+
+                return out
+
+            q = paddle.rand((1, 4, 2, 8)).astype("bfloat16")
+            _ = flash_attention_npu(q, q, q, 0.9, False, False)
+            paddle.nn.functional.flash_attention_npu = flash_attention_npu
+            return True
+        q = paddle.rand((1, 4, 2, 8)).astype("bfloat16")
+        _ = paddle.nn.functional.flash_attention.flash_attention(
+            q, q, q, 0.9, False, False
+        )
+        return True
+    except:
+        return False
+
+
+HAS_FLASH_ATTN = is_flash_attn_available()
+
+
+def has_flash_attn_func():
+    if HAS_FLASH_ATTN:
+        try:
+            if "npu" in paddle.get_device():
+                flash_attn_func_npu = paddle.nn.functional.flash_attention_npu
+                return flash_attn_func_npu, flash_attn_func_npu
+            else:
+                from paddle.nn.functional.flash_attention import (
+                    flash_attention as flash_attn_func,
+                )
+                from paddle.nn.functional.flash_attention import (
+                    flash_attn_unpadded as flash_attn_varlen_func,
+                )
+
+                return flash_attn_func, flash_attn_varlen_func
+        except:
+            return None, None
+    else:
+        return None, None

paddlex/inference/models/common/vlm/generation/__init__.py

@@ -0,0 +1,34 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration_utils import GenerationConfig
+from .logits_process import (
+    ForcedBOSTokenLogitsProcessor,
+    ForcedEOSTokenLogitsProcessor,
+    HammingDiversityLogitsProcessor,
+    LogitsProcessor,
+    LogitsProcessorList,
+    MinLengthLogitsProcessor,
+    RepetitionPenaltyLogitsProcessor,
+    TopKProcess,
+    TopPProcess,
+)
+from .stopping_criteria import (
+    MaxLengthCriteria,
+    MaxTimeCriteria,
+    StoppingCriteria,
+    StoppingCriteriaList,
+    validate_stopping_criteria,
+)
+from .utils import GenerationMixin