FlashAttention3 Forward Compatibility

This document is a quick reference for the current FlashAttention-3-compatible forward coverage provided by MATE on MUSA.

At a Glance

Area	Status	Notes
Q Mode	✅ Supported	Normal, Ragged, Padded
KV Mode	✅ Supported	Normal, Ragged, Padded, Paged
Append New KV	✅ Supported	flash_attn_with_kvcache appends via k / v; packed new KV is supported via cu_seqlens_k_new
RoPE Input	✅ Supported	flash_attn_with_kvcache supports rotary_cos / rotary_sin, Interleaved, and Non-interleaved
Cache Index Options	✅ Supported	cache_batch_idx, cache_leftpad
Mask Mode	✅ Supported	None, Causal, Local, Local + attention_chunk
Score Mode	✅ Supported	Standard softmax and softcap
Page Size	✅ Supported	1, 16, 64, and arbitrary page sizes
Dtype	✅ Supported	bf16, fp16, and torch.float8_e4m3fn forward inputs; FP8 uses q_descale / k_descale / v_descale scaling
HeadDim	✅ Supported	Any headdim <= 512
Optimization	✅ Supported	SplitKV, PackGQA, SchedulerMetadata
Output	✅ Supported	out, softmax_lse

MATE Extensions

Extension	Status	Notes
Context Parallel	✅ Supported	cp_world_size, cp_rank, cp_tot_seqused_k
Learnable Sink	✅ Supported	Supported on the local-attention path

Not Supported Yet

Feature	Status	Notes
FP8 + QV Input	❌ Not supported	FP8 forward inputs with qv are outside the current supported scope

Notes

• This page summarizes the compatibility surface, not every internal kernel detail.

• The statement Any headdim <= 512 refers to the supported forward-path head-dimension range.

• FP8 forward support currently refers to torch.float8_e4m3fn; pass optional q_descale, k_descale, and v_descale tensors with shape (batch_size, num_heads_kv) when scale factors are required.

• RoPE is supported only when appending new KV through k / v; rotary_dim must be <= headdim and divisible by 16.

• Local + attention_chunk requires MUSA SDK >= 5.1.0.

• For wrapper-level usage, see the FlashAttention wrapper page.