Conversation
Run VDN checkpoints through the existing H3 runner, encoders, VAEs, offload and Ulysses paths. Add hybrid window/linear attention and load branch weights with ordered default/turbo adapter merging. Key AdaLN caches by effective checkpoint and adapter identity (format 2). Add configurable noise generation and Qwen-only image references for matched benchmarks. Use deterministic frame reductions and contiguous Q activation inputs to resolve the observed repeatability and SP drift. Include the vdn_sp_8step.json profile, launch script and benchmark harness. Validation: - Ruff checks, Python/JSON syntax, shell syntax and staged whitespace checks. - Small-case full 8-step SP1/SP4 and independent SP4 repeats: 73/73 tensors exact. - Two full-size SP4 scenarios, each with 2 warmups and 3 measured requests; all five video, audio and MP4 hashes match within each scenario.
Remove duplicate validation, forwarding helpers, and redundant bookkeeping. Reuse read-only adapter readers and retain the existing cache identity and ordered LoRA merge behavior. Remove the VDN benchmark script. Validation: CPU attention, scheduler, weight/cache and communication checks; Ruff, JSON and shell checks; SP4 8-step regression with all 73 captured tensors exactly matching the previous implementation.
Reduce configure_vdn to reading model_spec.json and exposing the checkpoint path and attention configuration. Remove redundant artifact prechecks and its unused return value. Keep unsupported TP, quantization, and extra LoRA checks at model entry. Validation: released Stage-B/DMD configuration and adapter order match the previous version; all five model-entry conflict checks pass; Ruff and AST checks pass. No inference arithmetic changes.
Shard VDN attention heads and branch weights for TP and combined TP/SP. Merge ordered adapters before slicing weights and reuse H3 linear and offload components. Add optional tp_reproducible arithmetic with eight logical GEMM partitions, FP64 row reduction and seven-head VDN linear calls. TP1 uses the same partitions, so this mode establishes a new shared numerical baseline. Provide 8-step presets for TP8, SP8 and TP4+SP2 with an 8-GPU launcher. Validation: native TP1/TP8 at 768p, 124 frames, DMD8 and cached AdaLN produced identical 400 block outputs, 73 captured tensors and MP4 files. Focused CPU checks, Ruff, Python compilation, shell syntax and config checks passed.
Watebear
changed the title
Watebear
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Run VDN checkpoints through the existing H3 runner, encoders, VAEs, offload and Ulysses paths. Add hybrid window/linear attention and load branch weights with ordered default/turbo adapter merging.