Classes
class	SimulateRequest
class	CompareRequest

Functions
bool	_is_heavy_sim_model (dict[str, Any] cfg)
int	_sim_batch_limit (dict[str, Any] cfg)
list[torch.Tensor]	_grad_vector (torch.Tensor loss, list[torch.nn.Parameter] params, bool retain_graph=False)
list[torch.Tensor]	_hvp (torch.Tensor loss, list[torch.nn.Parameter] params, list[torch.Tensor] v)
list[torch.Tensor]	_hvp_from_fn (loss_fn, list[torch.nn.Parameter] params, list[torch.Tensor] v)
float	_grad_dot (list[torch.Tensor] ga, list[torch.Tensor] gb)
float	_tensor_list_norm (list[torch.Tensor] ts)
float	_normalized_influence (list[torch.Tensor] ga, list[torch.Tensor] gb)
list[torch.Tensor]	_lissa_inverse_hvp (test_loss_fn, train_loss_fn, int n_train, list[torch.nn.Parameter] params, float scale=25.0, float damping=0.05, int n_iter=10)
list[dict]	_influence_via_grad_dot (test_loss_fn, list[int] train_indices, list[dict] rows, list[torch.nn.Parameter] params, train_loss_fn)
dict[int, float]	_ntk_diagonal (torch.Tensor loss, list[torch.nn.Parameter] params, list[torch.Tensor] grads)
float	_power_iteration_max_eigenvalue (torch.Tensor loss, list[torch.nn.Parameter] params, int n_iter=3)
int	_hidden_state_index (model, int sae_layer)
torch.Tensor\|None	_sae_grad_scores_from_batch (model, sae, int sae_layer, torch.Tensor input_ids, torch.Tensor attention_mask, torch.Tensor\|None base_acts_mean)
dict[str, Any]	_run_dataset_quality (list[dict] rows)
dict	_detect_ai_generated (list[dict] rows)
None	_run_simulation (SimulateRequest req, asyncio.Queue queue, asyncio.AbstractEventLoop loop)
	training_simulate (SimulateRequest req)
Any	_sanitize (Any obj)
dict	compare_simulation_results (dict result_a, dict result_b, str label_a="Run A", str label_b="Run B", *, str run_id_a="", str run_id_b="")
	compare_simulations (CompareRequest req)

Variables
	router = APIRouter()
	DEVICE = resolve_compute_device()
	DTYPE = default_dtype_for_device(DEVICE)
int	_SIM_MAX_SAMPLES = 64
str	_SIM_FT_CKPT_PATH = None
str	_SIM_FT_MODEL_ID = None
list	_HARMFUL_KEYWORDS
list	_AI_FINGERPRINTS

Function Documentation

◆ _detect_ai_generated()

dict _detect_ai_generated ( list[dict] rows )

protected

Check if >30% of responses match common AI generation fingerprints.

Definition at line 445 of file train_simulate.py.

Referenced by _run_dataset_quality().

◆ _grad_dot()

float _grad_dot	(	list[torch.Tensor]	ga,
		list[torch.Tensor]	gb )

protected

Definition at line 167 of file train_simulate.py.

Referenced by _normalized_influence().

◆ _grad_vector()

list[torch.Tensor] _grad_vector	(	torch.Tensor	loss,
		list[torch.nn.Parameter]	params,
		bool	retain_graph = False )

protected

Detached per-parameter gradient vector for loss.

Definition at line 130 of file train_simulate.py.

Referenced by _influence_via_grad_dot(), _lissa_inverse_hvp(), and _run_simulation().

◆ _hidden_state_index()

int _hidden_state_index	(		model,
		int	sae_layer )

protected

Map SAE layer index → output_hidden_states tuple index.

Definition at line 284 of file train_simulate.py.

Referenced by _sae_grad_scores_from_batch().

◆ _hvp()

list[torch.Tensor] _hvp	(	torch.Tensor	loss,
		list[torch.nn.Parameter]	params,
		list[torch.Tensor]	v )

protected

Hessian-vector product H·v via double backprop.

Definition at line 142 of file train_simulate.py.

Referenced by _hvp_from_fn(), _ntk_diagonal(), and _power_iteration_max_eigenvalue().

◆ _hvp_from_fn()

list[torch.Tensor] _hvp_from_fn	(		loss_fn,
		list[torch.nn.Parameter]	params,
		list[torch.Tensor]	v )

protected

HVP with a fresh forward inside :  avoids reusing freed loss graphs.

Definition at line 156 of file train_simulate.py.

References _hvp().

Referenced by _lissa_inverse_hvp().

◆ _influence_via_grad_dot()

list[dict] _influence_via_grad_dot	(		test_loss_fn,
		list[int]	train_indices,
		list[dict]	rows,
		list[torch.nn.Parameter]	params,
			train_loss_fn )

protected

First-order influence: -∇L_test · ∇L_train (no Hessian). Reliable on PEFT/LoRA.

Definition at line 219 of file train_simulate.py.

References _grad_vector(), and _normalized_influence().

Referenced by _run_simulation().

◆ _is_heavy_sim_model()

bool _is_heavy_sim_model ( dict[str, Any] cfg )

protected

HF-only or wide models need lighter simulate passes (VRAM).

Definition at line 62 of file train_simulate.py.

Referenced by _run_simulation(), and _sim_batch_limit().

◆ _lissa_inverse_hvp()

list[torch.Tensor] _lissa_inverse_hvp	(		test_loss_fn,
			train_loss_fn,
		int	n_train,
		list[torch.nn.Parameter]	params,
		float	scale = 25.0,
		float	damping = 0.05,
		int	n_iter = 10 )

protected

Approximate H^{-1} · ∇L_test via LiSSA.
v_{t+1} = v_0 + (1 - damping)·v_t - (1/scale)·H·v_t
where v_0 = ∇L_test (held fixed).

train_loss_fn(j) must return a fresh scalar loss for batch j on each call.

Definition at line 192 of file train_simulate.py.

References _grad_vector(), and _hvp_from_fn().

Referenced by _run_simulation().

◆ _normalized_influence()

float _normalized_influence	(	list[torch.Tensor]	ga,
		list[torch.Tensor]	gb )

protected

Cosine-style influence in [-1, 1] :  stable for display and ranking.

Definition at line 181 of file train_simulate.py.

References _grad_dot(), and _tensor_list_norm().

Referenced by _influence_via_grad_dot(), and _run_simulation().

◆ _ntk_diagonal()

dict[int, float] _ntk_diagonal	(	torch.Tensor	loss,
		list[torch.nn.Parameter]	params,
		list[torch.Tensor]	grads )

protected

Per-parameter diagonal NTK entry via Rayleigh quotient:
  K_ii ≈ (g^T · H · g) / ||g||^2
Uses one HVP in the gradient direction.

Definition at line 241 of file train_simulate.py.

References _hvp().

Referenced by _run_simulation().

◆ _power_iteration_max_eigenvalue()

float _power_iteration_max_eigenvalue	(	torch.Tensor	loss,
		list[torch.nn.Parameter]	params,
		int	n_iter = 3 )

protected

Estimate the largest Hessian eigenvalue via power iteration.

Definition at line 262 of file train_simulate.py.

References _hvp().

Referenced by _run_simulation().

◆ _run_dataset_quality()

dict[str, Any] _run_dataset_quality ( list[dict] rows )

protected

Analyse dataset quality without loading the LLM.
Returns a quality report dict (no SSE streaming here :  caller handles that).

Definition at line 331 of file train_simulate.py.

References _detect_ai_generated().

Referenced by _run_simulation().

◆ _run_simulation()

None _run_simulation	(	SimulateRequest	req,
		asyncio.Queue	queue,
		asyncio.AbstractEventLoop	loop )

protected

Definition at line 460 of file train_simulate.py.

References _grad_vector(), _influence_via_grad_dot(), _is_heavy_sim_model(), _lissa_inverse_hvp(), _normalized_influence(), _ntk_diagonal(), _power_iteration_max_eigenvalue(), _run_dataset_quality(), _sae_grad_scores_from_batch(), _sim_batch_limit(), and _tensor_list_norm().

◆ _sae_grad_scores_from_batch()

torch.Tensor \| None _sae_grad_scores_from_batch	(		model,
			sae,
		int	sae_layer,
		torch.Tensor	input_ids,
		torch.Tensor	attention_mask,
		torch.Tensor \| None	base_acts_mean )

protected

Project loss gradient at the SAE layer through W_dec (works with PEFT/LoRA).

Definition at line 295 of file train_simulate.py.

References _hidden_state_index().

Referenced by _run_simulation().

◆ _sanitize()

Any _sanitize ( Any obj )

protected

Recursively replace NaN/Inf floats with None so JSON serialization never fails.

Definition at line 1278 of file train_simulate.py.

References _sanitize().

Referenced by _sanitize().

◆ _sim_batch_limit()

int _sim_batch_limit ( dict[str, Any] cfg )

protected

Definition at line 67 of file train_simulate.py.

References _is_heavy_sim_model().

Referenced by _run_simulation().

◆ _tensor_list_norm()

float _tensor_list_norm ( list[torch.Tensor] ts )

protected

Definition at line 174 of file train_simulate.py.

Referenced by _normalized_influence(), and _run_simulation().

◆ compare_simulation_results()

dict compare_simulation_results	(	dict	result_a,
		dict	result_b,
		str	label_a = "Run A",
		str	label_b = "Run B",
		*	,
		str	run_id_a = "",
		str	run_id_b = "" )

Diff two simulation payloads :  SAE features, influence, LR, attack-surface scores.

Definition at line 1289 of file train_simulate.py.

Referenced by compare_simulations().

◆ compare_simulations()

compare_simulations ( CompareRequest req )

Diff two simulation results. Returns structured comparison.

Definition at line 1500 of file train_simulate.py.

References compare_simulation_results().

◆ training_simulate()

training_simulate ( SimulateRequest req )

Streams simulation events as SSE. Client reads line-by-line.

Definition at line 1255 of file train_simulate.py.

Variable Documentation

◆ _AI_FINGERPRINTS

list aquin.compute.train_simulate._AI_FINGERPRINTS

protected

Initial value:

=  [
    "certainly!", "certainly,", "as an ai", "as an ai language model",
    "i'd be happy to", "i'd be happy to help", "i'm happy to help",
    "of course!", "of course,", "great question", "absolutely!",
    "sure, here", "sure! here", "i hope this helps", "feel free to ask",
    "i cannot and will not", "i want to emphasize", "it's important to note",
    "i need to be direct", "as a large language model",
]

Definition at line 118 of file train_simulate.py.

◆ _HARMFUL_KEYWORDS

list aquin.compute.train_simulate._HARMFUL_KEYWORDS

protected

Initial value:

=  [
    "bomb", "weapon", "kill", "murder", "hack", "malware", "ransomware",
    "phishing", "exploit", "suicide", "self-harm", "poison", "chemical weapon",
    "bioweapon", "synthesize drugs", "synthesize methamphetamine",
]

Definition at line 70 of file train_simulate.py.

◆ _SIM_FT_CKPT_PATH

str aquin.compute.train_simulate._SIM_FT_CKPT_PATH = None

protected

Definition at line 58 of file train_simulate.py.

◆ _SIM_FT_MODEL_ID

str aquin.compute.train_simulate._SIM_FT_MODEL_ID = None

protected

Definition at line 59 of file train_simulate.py.

◆ _SIM_MAX_SAMPLES

int aquin.compute.train_simulate._SIM_MAX_SAMPLES = 64

protected

Definition at line 57 of file train_simulate.py.

◆ DEVICE

aquin.compute.train_simulate.DEVICE = resolve_compute_device()

Definition at line 54 of file train_simulate.py.

◆ DTYPE

aquin.compute.train_simulate.DTYPE = default_dtype_for_device(DEVICE)

Definition at line 55 of file train_simulate.py.

◆ router

aquin.compute.train_simulate.router = APIRouter()

Definition at line 52 of file train_simulate.py.

Classes

Functions

Variables

Function Documentation

◆ _detect_ai_generated()

◆ _grad_dot()

◆ _grad_vector()

◆ _hidden_state_index()

◆ _hvp()

◆ _hvp_from_fn()

◆ _influence_via_grad_dot()

◆ _is_heavy_sim_model()

◆ _lissa_inverse_hvp()

◆ _normalized_influence()

◆ _ntk_diagonal()

◆ _power_iteration_max_eigenvalue()

◆ _run_dataset_quality()

◆ _run_simulation()

◆ _sae_grad_scores_from_batch()

◆ _sanitize()

◆ _sim_batch_limit()

◆ _tensor_list_norm()

◆ compare_simulation_results()

◆ compare_simulations()

◆ training_simulate()

Variable Documentation

◆ _AI_FINGERPRINTS

◆ _HARMFUL_KEYWORDS

◆ _SIM_FT_CKPT_PATH

◆ _SIM_FT_MODEL_ID

◆ _SIM_MAX_SAMPLES

◆ DEVICE

◆ DTYPE

◆ router