`nemo_automodel.components.models.deepseek_v4.kernels.sparse_attention`#

Autograd wrapper for vendored Miles DeepSeek V4 sparse-attention kernels.

Attribution:

Upstream project: Miles, https://github.com/yueming-yuan/miles
Upstream revision: e561465d0b9bbf06188b7a5e2020dc7fd691f732, deepseek-v4 branch
Upstream license: Apache-2.0, copyright 2025 Zhipu AI
Original source: https://github.com/yueming-yuan/miles/blob/e561465d0b9bbf06188b7a5e2020dc7fd691f732/miles_plugins/models/deepseek_v4/ops/attention_core.py

Module Contents#

`DeepSeekV4SparseAttention`	TileLang sparse MQA attention with custom backward.
`DeepSeekV4SparseAttentionHeadChunked`	TileLang sparse attention with smaller head groups and fp32 KV-grad accumulation.

`sparse_attn_tilelang`	Run vendored Miles DeepSeek V4 TileLang sparse attention.
`sparse_attn_tilelang_head_chunked`	Run vendored Miles sparse attention in TileLang head chunks.

class nemo_automodel.components.models.deepseek_v4.kernels.sparse_attention.DeepSeekV4SparseAttention#

Bases: torch.autograd.Function

TileLang sparse MQA attention with custom backward.

static forward( ctx, q: torch.Tensor, kv: torch.Tensor, attn_sink: torch.Tensor, topk_idxs: torch.Tensor, sm_scale: float | None = None, ) → torch.Tensor#: Run the vendored sparse attention forward kernel.

static backward( ctx, grad_output: torch.Tensor, ) → tuple[torch.Tensor, torch.Tensor, torch.Tensor, None, None]#: Run the vendored sparse attention backward kernel.

nemo_automodel.components.models.deepseek_v4.kernels.sparse_attention.sparse_attn_tilelang( q: torch.Tensor, kv: torch.Tensor, attn_sink: torch.Tensor, topk_idxs: torch.Tensor, sm_scale: float | None = None, ) → torch.Tensor#: Run vendored Miles DeepSeek V4 TileLang sparse attention.

class nemo_automodel.components.models.deepseek_v4.kernels.sparse_attention.DeepSeekV4SparseAttentionHeadChunked#

Bases: torch.autograd.Function

TileLang sparse attention with smaller head groups and fp32 KV-grad accumulation.

static forward( ctx, q: torch.Tensor, kv: torch.Tensor, attn_sink: torch.Tensor, topk_idxs: torch.Tensor, max_heads_per_kernel: int, sm_scale: float | None = None, ) → torch.Tensor#: Run the vendored sparse attention forward kernel over head chunks.

static backward( ctx, grad_output: torch.Tensor, ) → tuple[torch.Tensor, torch.Tensor, torch.Tensor, None, None, None]#: Run chunked backward and accumulate shared KV gradients in fp32.

nemo_automodel.components.models.deepseek_v4.kernels.sparse_attention.sparse_attn_tilelang_head_chunked( q: torch.Tensor, kv: torch.Tensor, attn_sink: torch.Tensor, topk_idxs: torch.Tensor, max_heads_per_kernel: int, sm_scale: float | None = None, ) → torch.Tensor#: Run vendored Miles sparse attention in TileLang head chunks.