Accumulate KV cache for chunk ``chunk_idx``. Two cache layouts are supported, distinguished by slot 6 (type flag): * **State-based**

Apply symmetric padding for bidirectional mode.

Apply causal padding using cached history.

Set numerically-neutral perf knobs (cuDNN bench, SDPA, Inductor).

Compiled output gate (fuses linear + silu + multiply).

Apply positional embedding to input tensor. Args: x: Input tensor (N, T, D) bs: Batch size Returns:

Compiled rotary embedding application (fuses view_as_complex + multiply chain).

Map compute capability → named arch bucket in `_CHUNKWISE_TUNING`. Blackwell (cap[0] >= 10) is split into "blackwell_dc" and "blackwell_spark"

Look up chunkwise kernel launch params from the tuning table. Resolution order: 1. `_ARCH_OVERRIDES[(cap, dot_prec)]` — exact-capability pi

Run backward (anti-causal) conv isolated per chunk. Within each chunk the input is time-flipped, the causal conv is applied, and the

Simulate non-causal conv by combining forward + backward causal passes. A causal depthwise Conv1d with kernel ``[w_0, w_1, ..., w_{k-1}]``

Build the per-block ``camctrl_cls`` schedule. The first ``camctrl_layers_num`` blocks use the camera-controlled attention. Within that range

Camera branch with cached delta-rule state.

Camera branch: SDPA with cached UCPE-transformed K, V.

Cached chunk-causal forward through the fused Triton scan. Same return contract as the torch ``CachedChunkCausalGDN.forward`` cached path (``

Short conv with cached left context: forward-cached + backward-isolated. Mirrors ``ChunkCausalGDN._apply_temporal_short_conv`` but replaces the

Run the camera-branch single-path delta-rule chunk-causal scan with two ``cam_scan_chunkwise`` calls (forward seeded from cached state + per-c

Streaming cam-branch QKV prep through the bidir's fused Triton kernel. Mirrors the prep section of :meth:`BidirectionalGDNUCPESinglePathLiteL

Run causal conv and preserve input dtype. Args: x: Tensor of shape (batch, seq_len, channels). conv: FLA ``ShortConvo

Clear possible gradient references in KV cache and cross-attention cache. This is important for preventing memory leaks, especially b

Compiled frame gate computation (fuses sigmoid + softplus + exp chain).

compute KL gradient (reference DMD implementation).

Compute residual connection for pixel unshuffle.

Compute residual connection.

Ensure downstream backward receives a contiguous gradient buffer.

Convert flow matching's prediction to x0 prediction. flow_pred: the prediction with shape [B, C, H, W] xt: the input noisy da

Constructs the noise schedule of Karras et al. (2022).

Bidirectional softmax camera branch (with UCPE transforms). Uses ``F.scaled_dot_product_attention`` with optional invalid-key masking.

Softmax attention (SDPA) reusing GDN parameters. Used by the hybrid GDN+Softmax architecture: every Nth block runs softmax attention instead

Chunk-causal softmax attention (SDPA / flex_attention) reusing GDN parameters. Used by ``ChunkCausalSoftmaxAttn``. Reuses ``qkv``, ``q_norm``,

Run the chunk-causal main GDN scan through the fused Triton chunkwise kernels. Uses two fused Triton calls (forward-with-state + per-chunk re

Generate a single chunk. Args: noise_chunk: noise input [batch_size, chunk_frames, C, H, W] chunk_start_fram

Frame-wise block-causal mask: full attention within each frame, causal across frames. Returns a boolean tensor of shape ``(1, 1, T*S, T*S)``

Return optimal kernel parameters for the current GPU. STATE_FP32: use fp32 state_prev when SRAM is large enough. - bf16 state_prev: ~96KB t

Randomly generate a timestep tensor based on the generator's task type. It uniformly samples a timestep from the range [min_timestep,

Handle empty temporal chunks by reserving cache slots.

Handle empty output when temporal dimension is 0.

Initialize camera short convs as identity to match base at step 0.

Initialize short conv as identity to match no-conv behavior at step 0.

initialize cached modules, refer to SelfForcingFlowEuler's implementation

Create empty 10-slot cache: ``kv_cache[chunk][block] = [None]*10``.