Add FA3 and tunable FlashInfer Solutions for GQA and MLA paged prefill
Adds three Solutions for the two paged-attention prefill Definitions, together
with the Implementation Configs that sweep their runtime knobs and the traces
that measure them. Both Definitions already exist; this contributes
implementations and measurements only.
What is added
| Solution | Definition | Backend |
|---|---|---|
fa3_gqa_prefill_v1 |
gqa_paged_prefill_causal_h16_kv1_d128_ps64 |
Flash Attention 3 |
flashinfer_gqa_prefill_tunable_v1 |
same | FlashInfer, knobs exposed |
flashinfer_mla_prefill_tunable_v1 |
mla_paged_prefill_causal_h16_ckv512_kpe64_ps1 |
FlashInfer, knobs exposed |
fa3_gqa_prefill_v1 is the first Flash Attention 3 Solution in the dataset. The
GQA prefill Definition previously had one Solution, so there was nothing to
compare against.
Tunable Solutions and their Implementation Configs
The existing Solutions fix their scheduling choices in the function body, so
there is nothing for an Implementation Config to sweep. Rather than change them,
these two variants take the same choices as keyword-only arguments:
| Solution | Runtime kwargs | Variants |
|---|---|---|
flashinfer_gqa_prefill_tunable_v1 |
backend, use_fp16_qk_reduction, disable_split_kv |
12 |
flashinfer_mla_prefill_tunable_v1 |
backend |
3 |
Defaults reproduce the monolithic baselines. The matrices contain only
combinations that were run against each Definition's own reference on H100 NVL
and matched at bf16 rounding scale — 12/12 for GQA, 3/3 for MLA. Nothing was
included on the strength of an API signature: cutlass and trtllm-gen are
accepted by the FlashInfer API but do not support these Definitions, so they are
excluded.
Notes on fa3_gqa_prefill_v1
FA3's flash_attn_with_kvcache expects a different KV layout than the Trace
provides, so the Solution bridges the two: it rebuilds the flat paged layout
(kv_indices + kv_indptr + kv_last_page_len) into the 2Dpage_table [batch, max_pages] and cache_seqlens [batch] FA3 wants, passes the
packed query through cu_seqlens_q since Q is stored ragged as[total_q, num_qo_heads, head_dim], applies the causal mask, and converts the
natural-log LSE FA3 returns to base 2 to match the FlashInfer convention the
Definition's reference uses.
Dependency declarations
spec.dependencies is derived from what each main.py actually imports.fa3_gqa_prefill_v1 declares flash_attn_interface, the top-level module the
FA3 hopper build installs (py_modules=["flash_attn_interface", ...]); there is
no flash_attn_3 Python package to name. As the first FA3 Solution here there
was no token to follow — happy to change it if you prefer another spelling.
Traces
Traces are appended to the existing per-Definition files; existing rows are
untouched.