file stringlengths 26 26 | meta dict | flash_out dict | k dict | q dict | v dict | md5 stringlengths 32 32 | bytes int64 619M 619M |
|---|---|---|---|---|---|---|---|
step00_block00.safetensors | {
"step": "0",
"source": "bf16 flash truth",
"block": "0",
"heads": "0-7",
"cfg_pass": "first (positive)",
"layout": "HND [8,S,D] bf16 (transposed from NHD for diag-format parity)"
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.4951
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.31377
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.25299
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 3.14422
} | 911c1dfe592982e56e06025bc1277fb2 | 619,315,720 |
step00_block33.safetensors | {
"step": "0",
"source": "bf16 flash truth",
"block": "33",
"heads": "0-7",
"cfg_pass": "first (positive)",
"layout": "HND [8,S,D] bf16 (transposed from NHD for diag-format parity)"
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.30829
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.16233
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.14679
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.79332
} | 8f21b6acef8f2b6b764d14e1120355e8 | 619,315,720 |
step00_block34.safetensors | {
"step": "0",
"source": "bf16 flash truth",
"block": "34",
"heads": "0-7",
"cfg_pass": "first (positive)",
"layout": "HND [8,S,D] bf16 (transposed from NHD for diag-format parity)"
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.25671
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.16455
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.19086
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.8619
} | e806787f4f9582706094049e8b1fb639 | 619,315,720 |
step00_block39.safetensors | {
"step": "0",
"source": "bf16 flash truth",
"block": "39",
"heads": "0-7",
"cfg_pass": "first (positive)",
"layout": "HND [8,S,D] bf16 (transposed from NHD for diag-format parity)"
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.27691
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.96674
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.01928
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.00615
} | cee5ab31ddc841081721d178049f284a | 619,315,720 |
step20_block00.safetensors | {
"step": "20",
"source": "bf16 flash truth",
"block": "0",
"heads": "0-7",
"cfg_pass": "first (positive)",
"layout": "HND [8,S,D] bf16 (transposed from NHD for diag-format parity)"
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.77676
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.33148
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.2577
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.68844
} | 555b2d570a33e6a5cec6e65508aed20d | 619,315,720 |
step20_block33.safetensors | {
"step": "20",
"source": "bf16 flash truth",
"block": "33",
"heads": "0-7",
"cfg_pass": "first (positive)",
"layout": "HND [8,S,D] bf16 (transposed from NHD for diag-format parity)"
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.5092
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.14236
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.07995
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.78755
} | 651a80c1e1f9bc80f260fb780a652864 | 619,315,720 |
step20_block34.safetensors | {
"step": "20",
"source": "bf16 flash truth",
"block": "34",
"heads": "0-7",
"cfg_pass": "first (positive)",
"layout": "HND [8,S,D] bf16 (transposed from NHD for diag-format parity)"
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.49197
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.29226
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.27208
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.92957
} | 874fd524b730a776fb4b917fcbc0b583 | 619,315,720 |
step20_block39.safetensors | {
"step": "20",
"source": "bf16 flash truth",
"block": "39",
"heads": "0-7",
"cfg_pass": "first (positive)",
"layout": "HND [8,S,D] bf16 (transposed from NHD for diag-format parity)"
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.31928
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.94384
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.06561
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.1185
} | c8e6a83aaeea36a78ccc379bc02d7d4b | 619,315,720 |
step39_block00.safetensors | {
"step": "39",
"source": "bf16 flash truth",
"block": "0",
"heads": "0-7",
"cfg_pass": "first (positive)",
"layout": "HND [8,S,D] bf16 (transposed from NHD for diag-format parity)"
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.91864
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.32397
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.22047
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.09049
} | 81127aec71dbd5e89354cdf02244b9b7 | 619,315,720 |
step39_block33.safetensors | {
"step": "39",
"source": "bf16 flash truth",
"block": "33",
"heads": "0-7",
"cfg_pass": "first (positive)",
"layout": "HND [8,S,D] bf16 (transposed from NHD for diag-format parity)"
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.62528
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.15148
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.17843
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.89696
} | 1166a3c1677eed9bd0181048f6a9f29f | 619,315,720 |
step39_block34.safetensors | {
"step": "39",
"source": "bf16 flash truth",
"block": "34",
"heads": "0-7",
"cfg_pass": "first (positive)",
"layout": "HND [8,S,D] bf16 (transposed from NHD for diag-format parity)"
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.57615
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.26819
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.27391
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.96509
} | cd93fb90de05aac315c64b140f9b73c7 | 619,315,720 |
step39_block39.safetensors | {
"step": "39",
"source": "bf16 flash truth",
"block": "39",
"heads": "0-7",
"cfg_pass": "first (positive)",
"layout": "HND [8,S,D] bf16 (transposed from NHD for diag-format parity)"
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.35219
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.0048
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 1.06891
} | {
"shape": [
8,
75600,
128
],
"finite": true,
"rms": 0.90546
} | b368490edc0e1cfc95d8756ae77c579f | 619,315,720 |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Wan2.2-T2V-A14B — BF16 flash-attention captures (truth set)
In-model QKV inputs and BF16 flash attention outputs captured from Wan2.2-T2V-A14B (81 frames, 40 steps, seed 42), taken during development of low-precision (MXFP4) block-sparse attention with nucleus (top-p) block routing. Intended for offline evaluation of block-sparse routing against full BF16 attention truth.
Files
12 × step{NN}_block{MM}.safetensors = 4 attention blocks {0, 33, 34, 39} × 3 denoise
steps {0, 20, 39} (these blocks are the forced-BF16-fallback / known-sensitive set
from the sparse work; steps span early / mid / late in the 40-step schedule).
Each file contains 4 tensors:
| tensor | shape | dtype | meaning |
|---|---|---|---|
q |
[8, 75600, 128] |
BF16 | query, HND (heads-first; transposed from the kernel's NHD layout) |
k |
[8, 75600, 128] |
BF16 | key |
v |
[8, 75600, 128] |
BF16 | value |
flash_out |
[8, 75600, 128] |
BF16 | attention output of the BF16 flash kernel on the same q/k/v — the truth |
Dimensions: 8 heads = heads 0–7, a deliberate subset of the model's 40 attention
heads (routing error is measured per query row and per head independently; the subset
keeps the set at ~7 GiB instead of ~35 GiB). S = 75600 = the full post-patchify
sequence for 81 frames at 1280×720 — latent grid 21×45×80 (temporal (81−1)/4+1=21,
spatial 90×160 latents patched 2×2), with no sequence sharding at the attention call.
128 = head dim (model inner dim 5120 = 40×128).
Positive CFG pass only (first pass of each step). All tensors verified finite;
per-file shapes/rms/md5 in captures_manifest.json.
Provenance
- Captured 2026-10-12 with an env-gated in-model hook in a vLLM-Omni-style Wan launch harness, single-card full sequence (no TP). Generation was the MXFP4-weight full-pipeline run; attention at the captured blocks was forced to BF16 flash (the ship "fallback strategy").
- Use case: replay nucleus routing (
keep = max(20% floor, smallest softmax-pool prefix ≥ top_p)at QPOOL 128/256/512) or fixed-k on these q/k/v and measure output error vsflash_outdirectly, without running the 454 s model.
Ship sparse configs these captures were made around: Wan QPOOL=256 / p=0.875 / floor 20% (model-level frame-cos vs dense 0.8865 mean, 1.16× speedup).
- Downloads last month
- 22