FlashKDA

FlashKDA: Flash Kimi Delta Attention — high-performance KDA kernels built on CUTLASS

News

2026-04-22 — Deep-Dive Blog: the design decisions behind FlashKDA v1, read it here.

Requirements

SM90 and above
CUDA 12.9 and above
PyTorch 2.4 and above

Installation

git clone https://github.com/MoonshotAI/FlashKDA.git flash-kda
cd flash-kda
git submodule update --init --recursive
pip install -v --no-build-isolation .

By default, the build detects the current CUDA device and compiles for that architecture. For wheel or CI builds, compile all supported architectures explicitly:

FLASH_KDA_CUDA_ARCHS=all pip install -v --no-build-isolation .

Supported values are auto (default), all, or a comma-separated arch list such as 90a,100a.

Using FlashKDA as an FLA backend

Once installed, FlashKDA is auto-dispatched from flash-linear-attention's chunk_kda. See fla-org/flash-linear-attention#852 for integration details.

Requirements

Install flash-linear-attention >= 0.5.0:
```
pip install -U flash-linear-attention
```

Call chunk_kda under torch.inference_mode()

import torch
from fla.ops.kda import chunk_kda

with torch.inference_mode():
    out, final_state = chunk_kda(
        q=q, k=k, v=v, g=g, beta=beta,
        scale=scale,
        initial_state=h0,
        output_final_state=True,
        use_gate_in_kernel=True,
        use_qk_l2norm_in_kernel=True,
        use_beta_sigmoid_in_kernel=True,
        safe_gate=True,
        A_log=A_log, dt_bias=dt_bias,
        lower_bound=lower_bound,
        transpose_state_layout=True,
        cu_seqlens=cu_seqlens,
    )

Opt out: set FLA_FLASH_KDA=0 to fall back to the Triton path.

Debug dispatch: add logging.basicConfig(level=logging.INFO) to see [FLA Backend] kda.chunk_kda -> flashkda on hit, or ... rejected: <reason> on miss.

Performance

See BENCHMARK_H20.md.

Tests

bash tests/test.sh

tests/test_fwd.py — correctness tests (exact match against the torch reference; compared with flash-linear-attention)

Kernel API

`flash_kda.fwd`

flash_kda.fwd(q, k, v, g, beta, scale, out, A_log, dt_bias, lower_bound,
              initial_state=None, final_state=None, cu_seqlens=None)

Parameters:

Parameter	Dtype	Shape	Description
`q`	bf16	`[B, T, H, K]`	Query
`k`	bf16	`[B, T, H, K]`	Key
`v`	bf16	`[B, T, H, V]`	Value
`g`	bf16	`[B, T, H, K]`	Gate before activation
`beta`	bf16	`[B, T, H]`	Beta logits (pre-activation; sigmoid applied internally)
`scale`	float	scalar	scaling factor
`out`	bf16	`[B, T, H, V]`	Output tensor
`A_log`	fp32	`[H]`	Log-gate parameter
`dt_bias`	fp32	`[H, K]`	Gate bias
`lower_bound`	float	scalar	Gate lower bound (range from -5.0 to 0)
`initial_state`	bf16/fp32/None	`[B, H, V, K]` or `[N, H, V, K]`	(optional) Initial recurrent state
`final_state`	bf16/fp32/None	`[B, H, V, K]` or `[N, H, V, K]`	(optional, output) Final recurrent state
`cu_seqlens`	int64	`[N+1]`	(optional) Cumulative sequence lengths for variable-length batching

Currently requires K = V = 128.
initial_state / final_state accept None (stateless), bf16, or fp32 tensors. When both are provided, their dtypes must match.
When cu_seqlens is provided, B must be 1, T is the total length across all sequences, and initial_state / final_state have shape [N, H, V, K].
When cu_seqlens is None, each batch element is treated as an independent sequence, and the state shape is [B, H, V, K].

Development

To set up IntelliSense (clangd) for the CUDA/C++ sources, run:

bash setup_clangd.sh

This generates a .clangd file with the correct repository paths and installs the global clangd config.yaml to ~/.config/clangd/.

Citation

@misc{flashkda2026,
      title={FlashKDA: Flash Kimi Delta Attention},
      author={Yutian Chen, Zhiyuan Li, Yucheng Wang, Ming Wei},
      year={2026},
      publisher = {GitHub},
      howpublished = {\url{https://github.com/MoonshotAI/FlashKDA}},
}