Jul 21 15:33 README: add novsmooth_linmant stress-probe row Dongyun Zou
15:33 stress probe: V-smooth OFF + 100% linmant (no collapse; p250 23.76 / p692 15.32 dB) Dongyun Zou
15:14 Update retimed 1+3 peak steady and locked benchmarks Dongyun Zou
14:55 Wan2.2 QK-INT8xPV-FP8 attention ablation 2026-07-21: results table + 15 sample videos Dongyun Zou
14:41 Add INT32 Direct FP8-Code Softmax report Dongyun Zou
13:47 Add V-smooth implementation and scheduling report Dongyun Zou
Document balanced-kmeans preprocessing, correction-warp reconstruction, row-sum LMMA, split-P retiming, optimization ablations, and LongCat/Wan measurements. Update archive index and SHA-256 manifest. 13:33 Add B200 LongCat 37k attention accuracy and speed report Dongyun Zou
Uploads the consolidated report, machine-readable summaries, final peak/steady/1350 MHz measurements, O1-O3 searches, GPU telemetry, validation artifacts, and Attention-QAT QK4/PV8 data. Jul 7 15:39 Update agenda in nunchux_low_bit_attention_blackwell_bundle.zip Dongyun Zou
15:39 Update agenda in nunchux_low_bit_attention_blackwell.html Dongyun Zou
15:39 Update agenda in nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
15:35 Upload self-contained slides zip bundle Dongyun Zou
15:32 Add B300 vsmooth on video: b300_vsmooth_on_A_dynamic_and_lively_00025.mp4 Dongyun Zou
15:32 Add B300 vsmooth on video: nunchux_low_bit_attention_blackwell.html Dongyun Zou
15:32 Add B300 vsmooth on video: nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
15:30 Embed B300 videos: b300_vsmooth_off_A_dynamic_and_lively_00025.mp4 Dongyun Zou
15:30 Embed B300 videos: ground_truth_A_dynamic_and_lively_00025.mp4 Dongyun Zou
15:30 Embed B300 videos: nunchux_low_bit_attention_blackwell.html Dongyun Zou
15:30 Embed B300 videos: nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
15:26 Update B300 visual comparison slide in nunchux_low_bit_attention_blackwell.html Dongyun Zou
15:26 Update B300 visual comparison slide in nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
15:21 Add B300 QK8 PV16 fake quant comparison to nunchux_low_bit_attention_blackwell.html Dongyun Zou
15:21 Add B300 QK8 PV16 fake quant comparison to nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
15:11 Add B200 v-kmeans vsmooth ablation results to nunchux_low_bit_attention_blackwell.html Dongyun Zou
15:11 Add B200 v-kmeans vsmooth ablation results to nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
15:09 Remove rel_l2 from slides in nunchux_low_bit_attention_blackwell.html Dongyun Zou
15:09 Remove rel_l2 from slides in nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
15:04 Add B300 section and blank B200 vsmooth table in nunchux_low_bit_attention_blackwell.html Dongyun Zou
15:04 Add B300 section and blank B200 vsmooth table in nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
14:52 Add B200 v-smooth ablation slides to nunchux_low_bit_attention_blackwell.html Dongyun Zou
14:52 Add B200 v-smooth ablation slides to nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
14:49 Add substitution step to approximation formula in nunchux_low_bit_attention_blackwell.html Dongyun Zou
14:49 Add substitution step to approximation formula in nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
14:46 Render formulas as HTML math in nunchux_low_bit_attention_blackwell.html Dongyun Zou
14:46 Render formulas as HTML math in nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
14:44 Compact math derivation slide for nunchux_low_bit_attention_blackwell.html Dongyun Zou
14:44 Compact math derivation slide for nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
14:40 Clarify variables in derivation slide for nunchux_low_bit_attention_blackwell.html Dongyun Zou
14:40 Clarify variables in derivation slide for nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
14:39 Update derivation slide in nunchux_low_bit_attention_blackwell.html Dongyun Zou
14:39 Update derivation slide in nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
14:34 Update simplified nunchux_low_bit_attention_blackwell.html Dongyun Zou
14:34 Update simplified nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
14:27 Upload nunchux_low_bit_attention_blackwell.html Dongyun Zou
14:27 Upload nunchux_low_bit_attention_blackwell.pdf Dongyun Zou
13:44 Upload ground_truth/A_dynamic_and_lively_00025.mp4 with huggingface_hub Dongyun Zou
13:43 Upload A_dynamic_and_lively_00025.mp4 with huggingface_hub Dongyun Zou
03:04 Render latency progress as continuous best-so-far curves Dongyun Zou
Remove dense benchmark scatter markers and connect observed best-so-far checkpoints with continuous lines. Keep a single dot only when an experiment has exactly one comparable result. 00:38 Refresh agent-token latency ablation plots Dongyun Zou
Regenerate all latency-vs-time and monotonic latency-vs-cumulative-agent-token plots from the latest benchmark ledgers and local session usage records. 00:37 Correct latency-vs-token plots to use agent session usage Dongyun Zou
Replace workload-token plots with monotonic best-so-far latency curves indexed by cumulative local agent-session token usage. Sources: Claude project JSONL, Codex token_count events, and OMH artifact usage. Add auditable token_curve_points.csv. 00:31 Add kernel ablation latency comparison plots Dongyun Zou
Upload GDN prefill, FP8 MoE, and DSA latency-vs-time and latency-vs-token plots, SVG sources, coverage metadata, and the reproducible plotting script. 
