Python class

TaylorSeer

`TaylorSeer`

class max.pipelines.modeling.base.TaylorSeer(max_order, dtype, device)

source

Bases: object

Standalone TaylorSeer caching module.

Compiles predict and update graphs at construction time and provides methods for the full TaylorSeer lifecycle: scheduling, prediction, factor updates, and state allocation.

Parameters:

max_order (int)
dtype (DType)
device (Device)

`compiled_predict()`

compiled_predict(factor_0, factor_1, factor_2, step_offset, max_order)

source

Run the compiled Taylor predict graph on eager tensors.

Parameters:

factor_0 (Tensor)
factor_1 (Tensor)
factor_2 (Tensor)
step_offset (Tensor)
max_order (Tensor)

Return type:

Tensor

`compiled_update()`

compiled_update(new_output, old_factor_0, old_factor_1, delta_step, max_order)

source

Run the compiled Taylor update graph on eager tensors.

Parameters:

new_output (Tensor)
old_factor_0 (Tensor)
old_factor_1 (Tensor)
delta_step (Tensor)
max_order (Tensor)

Return type:

tuple[Tensor, Tensor, Tensor]

`create_state()`

create_state(batch_size, seq_len, output_dim)

source

Allocate fresh per-request TaylorSeer state tensors.

Parameters:

batch_size (int)
seq_len (int)
output_dim (int)

Return type:

TaylorSeerState

`should_skip()`

static should_skip(step, warmup_steps, cache_interval)

source

Return True when the full transformer pass can be skipped at step.

Parameters:

step (int)
warmup_steps (int)
cache_interval (int)

Return type:

bool

TaylorSeer​

compiled_predict()​

compiled_update()​

create_state()​

should_skip()​

`TaylorSeer`

`compiled_predict()`

`compiled_update()`

`create_state()`

`should_skip()`