File size: 9,580 Bytes
e241361 | 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 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 | # Bernini Editing 下一步优化规划
更新日期:2026-07-02
当前同节点五例 hot-inference 结果为 pristine `1714.106 s`、editing lossless
`1138.782 s`,总和口径 `1.505x`。后续工作分成 **lossless system track** 和
**approximate algorithm track**,两条轨道使用不同正确性和质量门槛,结果不得混报。
## 统一定义
### Lossless system track
保持同一实数算法和数学映射:模型权重、40-step scheduler、dense attention、
CFG/APG 分支、conditioning 和随机采样不变。允许等价重排、批处理、通信 overlap、
cache 和不同的 BF16 reduction tiling,因此不要求跨实现 bitwise equality。
通过条件是公式/排列 oracle、serial differential、RNG 与 cache lifecycle test、真实
GPU shape audit 和五例同节点 hot A/B 全部通过。PSNR/SSIM、SHA 和人工 contact sheet
只作诊断,不代替数学语义审计。
### Approximate algorithm track
PISA sparse attention、FP8 quantization、EasyCache、TeaCache、step skipping 和其他
改变算子或实际计算量的方法都属于近似轨道。必须同时报告速度、真实执行工作量、
数值误差和视频质量;即使视觉结果可接受,也不能标为 lossless。
## P0:强化基准与审计
1. 在同一 allocation 中交替运行 pristine 与 optimized,至少三轮 measure,报告每例
median、P90、range 和 `sum(pristine)/sum(optimized)`,降低节点时序噪声。
2. 增加 request-stage timer:media preprocess、planner、T5、expert H2D/switch、
DiT normal step、VAE decode、D2H/postprocess。所有 stage 合计必须闭合到 synchronized
request walltime。
3. 固化 Torch/CUDA/FA3/NCCL/Triton、model snapshot、输入 SHA、launcher env、GPU
clocks 和峰值显存;统一 pristine/optimized 的 FA3 wrapper 后再复测 dependency
normalization 的影响。
4. 补一个 live 8-GPU `all_to_all_single` vs legacy `all_to_all` differential,覆盖五例
实际 source+target sequence length、global padding trim 和 async stream event。
5. 扩大 real-Wan differential 到 expert 1/2、多个 block 和真实五例 metadata;逐层
检查 packed segment 不串扰,而不是追求最终视频逐比特一致。
## P1:Lossless 条件路径去重
### Rank-leader planner/encoder
当前八个 Ulysses rank 重复执行 media condition、Qwen planner 和 T5。评估由 group
leader 计算确定性 embedding/distribution parameters,再广播最终 tensor 和 metadata。
必须保持 dtype、shape、stride、mask、token ordering 与原路径一致;涉及 latent sample
时只广播 distribution parameters,仍在原逻辑位置消费 RNG,或证明一次采样广播与
原分布/RNG 合同等价。
验证顺序:单模块 broadcast microbench、两例 2-step differential、五例 40-step
同节点 A/B。只有 E2E 有稳定正收益且 69/69 CPU suite、GPU formula audit 和 cache
lifecycle 均通过,才进入默认 launcher。
### 跨请求 media cache
把当前 request-local media dedupe 扩展为 server-lifetime immutable cache,key 至少包含
文件内容 SHA、duration/fps、resize/crop/normalize 参数、model snapshot、dtype 和
encoder version。只缓存确定性 payload 或 distribution parameters,不缓存本应重新
sample 的 latent;提供 LRU、显存/主存预算和请求结束引用释放。
### Condition/metadata cache 整理
统一 condition projection、RoPE、mask、`cu_seqlens` 和 packed-condition 的 key 与
失效策略。为 tensor mutation/version、expert switch、shape collision、异常 teardown、
并发 request 隔离增加测试,避免为了命中率放宽语义条件。
## P1:Lossless 权重搬运与 overlap
### Root H2D + NCCL broadcast
当前每个 rank 从 host 各自搬运同一 expert。按真实 QKV/out、FFN-up/down 和整 expert
bucket 比较:all-rank H2D、rank-0 H2D + NCCL broadcast、分片 H2D + all-gather。
microbenchmark 必须计入 pinning、allocator、event 和 collective,同一 weight bit
pattern 到达每个 rank 后再做 module output differential。
### Expert 2 延迟预取
在 expert 1 的最后若干 step 用独立 copy stream 预取 expert 2,并用 event 精确约束
首次消费。不得恢复双 expert 全程 resident,因为已有 profiling 显示它会拖慢 normal
step。验收同时看 request E2E、normal-step P50/P90 和 peak memory。
### Offload 生命周期
继续审计 `CpuMasterModule` 的 parameter replacement、shared storage、异常退出和 repeated
request teardown。探索 pinned CPU master 与 bucketed async H2D,但禁止在 timer 外偷跑
属于当前请求的必要 transfer。
## P2:Lossless DiT 与 decode
1. 对固定 shape bucket 评估 `torch.compile` 和 CUDA Graph。expert switch、不同 editing
layout、packed branch count 和 cache hit/miss 应拆成显式 graph;先验证 collective
次序和 RNG state,再测 E2E。
2. 继续融合 target patch、mask gather、APG delta 和 output unpatchify 中的纯 elementwise
链。每个 fused kernel 都对独立 FP32/FP64 公式 oracle,并使用实际 5120 hidden dim、
BF16 和 production token length。
3. 对 RV2V 动态 branch set 自动选择 chunk 1-5,目标是减少 launch overhead 与峰值显存,
但 branch boundary、zero-scale alias 和 APG reference 必须与 serial oracle 一致。
4. 拆分 VAE decode、D2H、layout/color conversion 和 MP4 encode。热 latency 只可 overlap
真正能与同请求前序并行的工作;后台未完成任务不能从 latency 中删除。
5. 评估 channels-last、decode-only compile、pinned async D2H 和预分配 output buffer,
逐项验证 decode tensor 的数值容差和 frame/layout 顺序。
## P3:Approximate attention 与低精度
### PISA
先按 editing 的真实 module shape 做 self-attention microbenchmark,完整计入 routing、
top-k、layout conversion、sparse kernel 和 concat。扫描 density、layer range、timestep
schedule 与 V2V/RV2V branch shape;只有模块端到端正收益后才集成完整五例。
PISA 报告必须包括实际 attention density、每层/每 timestep 命中、E2E/stage speedup、
数值误差、PSNR/SSIM、文本一致性、运动指标和人工盲评。cross-attention 是否保持 dense
要单独写明。
### FP8
在模型加载时预生成 FP8 weight/scale,避免 request 内 lazy quant。分别测 QKV/out、
FFN-up/down、cross-KV 的真实 M/K/N,比较 per-tensor、per-channel 和 timestep-aware
activation scale。module gate 需要同时报告含 quant/layout 的速度、误差和显存;FP8
始终属于 approximate track。
## P3:Approximate diffusion cache
### EasyCache
对 V2V、MV2V、RV2V、ADS2V 分任务记录 high/low-noise expert 的实际 DiT forward 数和
hit schedule。扫描 threshold、retain-first、retain-last 和最小执行预算;速度必须从
同运行、只关闭 EasyCache 的 optimized-stack baseline 计算。
### TeaCache
不能直接复用 Wan2.1/T2V polynomial。先用 Bernini editing 双 expert 的 residual curve
分别拟合系数,再在与 EasyCache 相同的实际 forward budget 下比较最差 case、temporal
flicker 和 source/reference adherence。
### 组合档
PISA、FP8 与 diffusion cache 会相互改变 residual 和数值轨迹。组合实验必须重新校准,
不能把单项 speedup 相乘,也不能沿用单项的 executed-step 或质量数据。推荐档至少保留:
- dense lossless reference;
- one-method optimized-stack baseline;
- PISA-only、cache-only、FP8-only;
- 明确参数的组合候选。
## 统一实验门槛
每个候选按以下顺序推进:
1. 写明数学变换、可复用条件、潜在失效模式和理论上限。
2. 用实际 dtype/shape/layout 做 module microbenchmark,报告 warmup、P50/P90、峰值显存
和所有前后处理开销。
3. 运行 CPU semantic suite、2-step 单例和相关 GPU oracle。
4. 同节点跑五例 warmup+measure;lossless 候选保持 dense 40-step,approximate 候选记录
实际 DiT forward、density 或 quantization 配置。
5. lossless 以算法/数学语义通过为 gate;approximate 另做逐帧指标、source adherence、
motion/text metrics 和人工评价。两者的表格、命名和结论分开。
6. 只保留同节点重复实验中的稳定正收益;负结果和 no-op 也记录,避免重复探索。
## 近期交付顺序
| 优先级 | 交付物 | 目标 |
| --- | --- | --- |
| P0 | 三轮同 allocation 五例 A/B + stage ledger | 收紧 1.505x 的方差和瓶颈解释 |
| P0 | live 8-GPU collective + real metadata audit | 补齐当前 world-size-2 硬件覆盖限制 |
| P1 | rank-leader planner/T5 prototype | 删除 Ulysses rank 间确定性重复计算 |
| P1 | root-H2D/broadcast microbench | 降低每 rank 重复 expert transfer |
| P1 | expert-2 prefetch prototype | 隐藏不可避免的 expert switch H2D |
| P1 | server-lifetime immutable media cache | 降低重复素材请求的 preprocess 成本 |
| P2 | graph/compile + decode microbench | 降低固定 launch 和 decode 开销 |
| P3 | PISA/FP8 module gates | 决定近似 kernel 是否值得 full integration |
| P3 | Editing-specific EasyCache/TeaCache calibration | 建立独立、可解释的速度-质量档位 |
所有新速度必须同时注明分母、commit、job、节点、计时边界和是否包含 model load。
`pristine baseline` 只作 upstream 锚点;候选的增量收益必须相对同运行、仅关闭该候选的
`optimized-stack baseline`。
|