| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
feat(phyai): add Dreamzero support (#65) * feat(phyai): add DreamZero support on latest main * refactor(dreamzero): reuse framework layers and runtime TF32 settings * feat(phyai): add policy-selectable L2 RMSNorm kernel * fix(phyai): validate L2 RMSNorm and migrate text linear layers * fix(dreamzero): reuse compiled UniPC updates across timesteps Pass current and historical sigma tensors into compiled updates so complete schedules reuse bounded order variants. Preserve the original GPU BF16 trajectory. * fix(dreamzero): preserve input shapes and guidance options Normalize vector state/action samples before padding, propagate explicit negative prompts through both processor steps, and skip negative encoding and broadcasts when effective guidance disables CFG. * fix(dreamzero): correct VAE batches and encoder offload Allocate tiled decoder outputs for the full batch, convert encoder tile units once per call, and construct gated image-encoder projections on CPU before loading and activation. * refactor(dreamzero): simplify inference and add CuTe T5 GELU Remove legacy WS1 and sequential CPU offload interfaces, redundant cache wrappers, and unused encoder paths. Reuse shared tensor-parallel layers and select FlashInfer's FA2 backend for DreamZero attention. Add a JIT CUTLASS CuTe T5 GELU kernel for contiguous and strided inputs, with corresponding layer/kernel tests and benchmarks. Preserve strict checkpoint loading, numerical precision, and cache lifecycle contracts. Validation: targeted CUDA kernel/layer checks, 43 final CUDA audit cases including two-rank TP/CFG and fake checkpoint coverage, and scripts/run_pre_commit.sh. * docs(skills): generalize PR review and validation guidance Require revision-aware feedback checks, shared-layer contract verification, and complete removal of retired interfaces. Cover cache ownership, device defaults, tensor parallelism, numerical equivalence, and kernel benchmarks. Keep PR-specific evidence in work logs and preserve the user's authorization boundaries when fixing code or interacting with GitHub. * docs(dreamzero): add bilingual inference and accuracy guides * docs(dreamzero): address final review feedback --------- Co-authored-by: lywbarca <lywbarca@users.noreply.github.com> Co-authored-by: chenghuaWang <chenghua.wang.edu@gmail.com> | 5 天前 | |
feat(phyai): add Dreamzero support (#65) * feat(phyai): add DreamZero support on latest main * refactor(dreamzero): reuse framework layers and runtime TF32 settings * feat(phyai): add policy-selectable L2 RMSNorm kernel * fix(phyai): validate L2 RMSNorm and migrate text linear layers * fix(dreamzero): reuse compiled UniPC updates across timesteps Pass current and historical sigma tensors into compiled updates so complete schedules reuse bounded order variants. Preserve the original GPU BF16 trajectory. * fix(dreamzero): preserve input shapes and guidance options Normalize vector state/action samples before padding, propagate explicit negative prompts through both processor steps, and skip negative encoding and broadcasts when effective guidance disables CFG. * fix(dreamzero): correct VAE batches and encoder offload Allocate tiled decoder outputs for the full batch, convert encoder tile units once per call, and construct gated image-encoder projections on CPU before loading and activation. * refactor(dreamzero): simplify inference and add CuTe T5 GELU Remove legacy WS1 and sequential CPU offload interfaces, redundant cache wrappers, and unused encoder paths. Reuse shared tensor-parallel layers and select FlashInfer's FA2 backend for DreamZero attention. Add a JIT CUTLASS CuTe T5 GELU kernel for contiguous and strided inputs, with corresponding layer/kernel tests and benchmarks. Preserve strict checkpoint loading, numerical precision, and cache lifecycle contracts. Validation: targeted CUDA kernel/layer checks, 43 final CUDA audit cases including two-rank TP/CFG and fake checkpoint coverage, and scripts/run_pre_commit.sh. * docs(skills): generalize PR review and validation guidance Require revision-aware feedback checks, shared-layer contract verification, and complete removal of retired interfaces. Cover cache ownership, device defaults, tensor parallelism, numerical equivalence, and kernel benchmarks. Keep PR-specific evidence in work logs and preserve the user's authorization boundaries when fixing code or interacting with GitHub. * docs(dreamzero): add bilingual inference and accuracy guides * docs(dreamzero): address final review feedback --------- Co-authored-by: lywbarca <lywbarca@users.noreply.github.com> Co-authored-by: chenghuaWang <chenghua.wang.edu@gmail.com> | 5 天前 |