File size: 2,243 Bytes
80692f2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 | name: 02_kda_cutlass
display_name: "Kimi Delta Attention (chunk forward) via CUTLASS CuTe"
precision: bf16
regime: compute
# Approximate FLOPs per forward call. KDA forward (chunk form) breakdown per
# token across all heads:
# - intra-chunk Aqk (q @ k.T with decay): 2*BT*K
# - intra-chunk Akk + cumulative invert : ~3*BT*K + BT^2
# - w = A @ (g.exp()*k): 2*BT*K
# - u = A @ v: 2*BT*V
# - inter-chunk q @ S: 2*K*V
# - inter-chunk (Aqk @ v_i): 2*BT*V
# - state update S += k.T @ v_i: 2*K*V
# Aggregating with a constant of 4 (typical for chunked linear attention) gives
# the formula below. Used only for roofline reporting -- not graded directly.
flops_formula: "4 * B * T * H * (K*V + CHUNK_SIZE*K + CHUNK_SIZE*V)"
# Bytes moved (lower bound): one read of q,k,v,g,beta and one write of o.
bytes_formula: "B*T*H*K*2 + B*T*H*K*2 + B*T*H*V*2 + B*T*H*K*4 + B*T*H*2 + B*T*H*V*2"
hardware: [RTX_PRO_6000]
peak_tflops_key: bf16 # KDA's heavy work is bf16 @ bf16 -> fp32 GEMM-equivalent
peak_bandwidth_key: dram
# bf16 default tolerance from src/eval/correctness.py is 1e-2 atol/rtol; KDA's
# long recurrence accumulates more error than a single GEMM, so we relax slightly.
tolerance:
torch.bfloat16: 0.05 # was "5e-2" -- PyYAML parses un-dotted sci-notation as string
# Forbidden ops -- using any in solution.py fails correctness post-hoc.
# The whole point of the problem is to write the kernel; calling FLA's existing
# Triton chunk_kda is the canonical reward-hack to block.
forbidden:
- "fla.ops.kda"
- "fla.ops.chunk_kda"
- "chunk_kda"
- "fused_recurrent_kda"
- "naive_chunk_kda"
- "naive_recurrent_kda"
sota:
name: "FLA chunk_kda (Triton)"
url: "https://github.com/fla-org/flash-linear-attention/tree/main/fla/ops/kda"
function: "fla.ops.kda.chunk_kda"
deps:
- "flash-linear-attention>=0.3"
# Documented H100 throughput (informational, not graded). FLA's KDA Triton
# kernel hits roughly 0.6-0.8x of FlashAttention-2 wall-clock on H100 at the
# B=2,T=2048,H=8,K=V=128 shape (per the Kimi Linear blog / FLA benchmarks).
reference_throughput_tflops_h100: null
num_correct_trials: 3
num_perf_trials: 20
|