flashinfer.cake_dsa_indexer_nondet_train¶
Cake DSA indexer top-k for training (32 heads, head dimension 128, BF16 q
/ k, FP32 w) for SM100a, SM103a and SM107a: for every query, the exact
best min(top_k, visible) keys by the indexer score, ties to the larger id.
The module is a stable API; its backend package,
flashinfer.experimental.cake.cake_dsa_indexer_nondet_train, is
experimental and is imported on first use.
The mode is part of every name, and no entry takes a deterministic
keyword. Per row, the set of selected ids and score bits repeats bit for bit,
and with it the number of valid slots. The slot order within a row and the
positions of the -1 / -inf padding can change between calls. The
deterministic package flashinfer.cake_dsa_indexer_det_train selects the
same set and returns it in ascending id order. Under
torch.use_deterministic_algorithms(True) a call with queries and keys
raises PyTorch’s alert before it allocates or launches anything; with
warn_only=True it warns and runs. It never falls back to a deterministic
implementation. This package offers the BF16 form only.
|
Fused DSA indexer scoring and exact top-k selection with set-equality repeatability (DeepSeek Sparse Attention indexer, training), SM100 / SM103 / SM107 only. |
Explicit workspace bound of |