ml.models.architectures.rwkv

Defines modules for RWKV blocks.

RWKV blocks are similar to Transformer blocks, but use a different attention mechanism that doesn’t require a linearly growing KV cache.

Training requires CUDA kernel requires installing triton:

pip install triton

class ml.models.architectures.rwkv.WkvWithEps(*args, **kwargs)

Bases: Function

static forward(ctx: FunctionCtx, w: Tensor, u: Tensor, k: Tensor, v: Tensor, state: Tensor) → tuple[torch.Tensor, torch.Tensor]

This function is to be overridden by all subclasses. There are two ways to define forward:

Usage 1 (Combined forward and ctx):

@staticmethod
def forward(ctx: Any, *args: Any, **kwargs: Any) -> Any:
    pass

• It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

• See combining-forward-context for more details

Usage 2 (Separate forward and ctx):

@staticmethod
def forward(*args: Any, **kwargs: Any) -> Any:
    pass

@staticmethod
def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
    pass

• The forward no longer accepts a ctx argument.

• Instead, you must also override the torch.autograd.Function.setup_context() staticmethod to handle setting up the ctx object. output is the output of the forward, inputs are a Tuple of inputs to the forward.

• See extending-autograd for more details

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

static backward(ctx: FunctionCtx, grad_wkv: Tensor, grad_state: Tensor) → tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]

Defines a formula for differentiating the operation with backward mode automatic differentiation (alias to the vjp function).

This function is to be overridden by all subclasses.

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

ml.models.architectures.rwkv.initial_state_with_eps(emb_dim: int) → Tensor

ml.models.architectures.rwkv.wkv_with_eps(w: Tensor, u: Tensor, k: Tensor, v: Tensor, state: Tensor) → tuple[torch.Tensor, torch.Tensor]

Runs the core WKV computation.

Parameters:

• w – The decay tensor, with shape (D)

• u – The output multiplier tensor, with shape (D)

• k – The K tensor, with shape (B, T, D)

• v – The V tensor, with shape (B, T, D)

• state – The state tensor, with shape (B, 3, T, D), consisting of the alpha, beta and eps tensors, each with shape (B, 1, T, D)

Returns:

The WKV tensor, with shape (B, T, D), and the next state, with shape (B, 3, 1, D), consisting of the next alpha, beta and eps tensors, each with shape (B, 1, 1, D)

class ml.models.architectures.rwkv.WkvLogSpace(*args, **kwargs)

Bases: Function

static forward(ctx: FunctionCtx, w: Tensor, u: Tensor, k: Tensor, v: Tensor, state: Tensor) → tuple[torch.Tensor, torch.Tensor]

This function is to be overridden by all subclasses. There are two ways to define forward:

Usage 1 (Combined forward and ctx):

@staticmethod
def forward(ctx: Any, *args: Any, **kwargs: Any) -> Any:
    pass

• It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

• See combining-forward-context for more details

Usage 2 (Separate forward and ctx):

@staticmethod
def forward(*args: Any, **kwargs: Any) -> Any:
    pass

@staticmethod
def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
    pass

• The forward no longer accepts a ctx argument.

• Instead, you must also override the torch.autograd.Function.setup_context() staticmethod to handle setting up the ctx object. output is the output of the forward, inputs are a Tuple of inputs to the forward.

• See extending-autograd for more details

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

static backward(ctx: FunctionCtx, grad_wkv: Tensor, grad_state: Tensor) → tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]

Defines a formula for differentiating the operation with backward mode automatic differentiation (alias to the vjp function).

This function is to be overridden by all subclasses.

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

ml.models.architectures.rwkv.initial_state_log_space(emb_dim: int) → Tensor

ml.models.architectures.rwkv.wkv_log_space(w: Tensor, u: Tensor, k: Tensor, v: Tensor, state: Tensor) → tuple[torch.Tensor, torch.Tensor]

Runs the core WKV computation.

Parameters:

• w – The decay tensor, with shape (D)

• u – The output multiplier tensor, with shape (D)

• k – The K tensor, with shape (B, T, D)

• v – The V tensor, with shape (B, T, D)

• state – The state tensor, with shape (B, 3, D), consisting of the alpha plus, alpha minus and beta tensors, each with shape (B, 1, D)

Returns:

The WKV tensor, with shape (B, T, D), and the next state, with shape (B, 2, D), consisting of the next alpha plus, alpha minus and beta tensors, each with shape (B, 1, D)

ml.models.architectures.rwkv.get_wkv_fn(key: Literal['eps', 'log']) → Callable[[Tensor, Tensor, Tensor, Tensor, Tensor], tuple[torch.Tensor, torch.Tensor]]

ml.models.architectures.rwkv.get_wkv_fn_cuda(key: Literal['eps', 'log']) → Callable[[Tensor, Tensor, Tensor, Tensor, Tensor], tuple[torch.Tensor, torch.Tensor]]

ml.models.architectures.rwkv.get_default_wkv_fn_key() → Literal['eps', 'log']

class ml.models.architectures.rwkv.RwkvAttention(dim: int, wkv_key: Literal['eps', 'log'] | None = None)

Bases: Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

init_x: Tensor

init_state: Tensor

time_shift(last_x: Tensor, x: Tensor) → Tensor

forward(x: Tensor, state: tuple[torch.Tensor, torch.Tensor] | None) → tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class ml.models.architectures.rwkv.RwkvFeedForward(dim: int, ffn_dim: int)

Bases: Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

init_state: Tensor

time_shift(last_x: Tensor, x: Tensor) → Tensor

forward(x: Tensor, state: Tensor | None = None) → tuple[torch.Tensor, torch.Tensor]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class ml.models.architectures.rwkv.RwkvBlock(emb_dim: int, pre_norm: bool, wkv_key: Literal['eps', 'log'] | None = None)

Bases: Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

run_attn(x: Tensor, state: tuple[tuple[torch.Tensor, torch.Tensor], torch.Tensor] | None = None) → tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]

run_ffn(x: Tensor, state: tuple[tuple[torch.Tensor, torch.Tensor], torch.Tensor] | None = None) → tuple[torch.Tensor, torch.Tensor]

forward(x: Tensor, state: tuple[tuple[torch.Tensor, torch.Tensor], torch.Tensor] | None = None) → tuple[torch.Tensor, tuple[tuple[torch.Tensor, torch.Tensor], torch.Tensor]]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class ml.models.architectures.rwkv.RwkvStack(emb_dim: int, num_layers: int, wkv_key: Literal['eps', 'log'] | None = None)

Bases: Module

Defines a stack of RWKV modules.

Parameters:

• emb_dim – The number of embedding dimensions in each block

• num_layers – The number of layers in the stack

• wkv_key – The WKV algorithm to use

Inputs:

x: The input tensor, with shape (B, T, D) state: The previous state

Outputs:

The output tensor, with shape (B, T, D), and the next state

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor, state: list[tuple[tuple[torch.Tensor, torch.Tensor], torch.Tensor]] | None = None) → tuple[torch.Tensor, list[tuple[tuple[torch.Tensor, torch.Tensor], torch.Tensor]]]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.