Minor build fixes#2
Merged
Merged
Conversation
Green-Sky
approved these changes
Aug 13, 2023
Closed
idostyle
added a commit
to idostyle/stable-diffusion.cpp
that referenced
this pull request
Jan 11, 2025
idostyle
added a commit
to idostyle/stable-diffusion.cpp
that referenced
this pull request
May 13, 2025
dbrain
added a commit
to dbrain/hbd-longcat-avatar.cpp
that referenced
this pull request
Jun 6, 2026
…O-GO) Add WanVAE::encode_partial (per-chunk streaming primitive, mirror of decode_partial) + enc_feat_idx:* encoder-cache branch in build_graph_partial, and WanVAE::encode_tail — an in-graph streaming encode (single forward graph) that primes the causal encoder cache with prior tail frames then encodes the target as a cached chunk. encode_tail is the working path: the cross-graph build_graph_partial cache is disabled/buggy for chunk>=1, and it front-pads the priming tail to size 1+4k so the encoder only sees chunk sizes 1 or 4. WanVAERunner::encode_streaming wraps it (with the [0,1]->[-1,1] input scaling the encode() wrapper does); nava encode-stream exposes it as a tool. Gate result (HANDOFF-nava-STREAMING-VAE-ENCODE-GATE.md): the cached encode injects a real tail-dependent change but it is direction-blind (forward vs reversed motion alias to cos 0.953), stays I-like in std, and is OOD as a standalone anchor. A single VAE latent frame cannot carry motion direction -> frozen-model single-frame continuity is capped. Recommend the LoRA continuation adapter.
dbrain
added a commit
to dbrain/hbd-longcat-avatar.cpp
that referenced
this pull request
Jun 11, 2026
LTX-2.3 joint audio-video long-form work (eval, flag-gated; avatar/t2v byte-identical when new paths unused). leejet#1 VRAM — AdaLN modulation-collapse port (98e8d16): per-token video timesteps carry few UNIQUE values; compute modulation on the compact uniq set and ggml_get_rows-gather back per token. New gather_mod_tokens helper + optional expand_sel on get_ada_values / get_output_scale_shift; host-side collapse builds the int32 selector in build_graph (ltx_video_token_sel). −62% DiT activation VRAM, bit-exact. leejet#2 In-process N-segment chaining: resident-DiT chain loop in cli (DiT loaded once, DiT_loads=1), in-memory cont_latent (5D cont_tail), per-segment --ltx-chain-prompts / --ltx-chain-audio-dir, absolute-timeline audio slicing. Two resident-free GGML_ASSERT(buffer) guards added (audio VAE + gemma TE: gate frees on !keep_diffusion_model_resident). High-res NaN fix (root cause, not band-aid): F16 K/V cast overflow in LTXAV CrossAttention self-attn → one inf, then unmasked softmax spreads it to 100% NaN. Fix = kv_scale=1/256 on the cross-attn path (env override LTX_ATTN_KV_SCALE). Exact repro 100%-NaN→0; full 1280x704 / 8-beat / 8-step movie renders the whole chain clean. Instrumentation: LTX_NAN_DEBUG (per-step in/out nan/inf/range scan), LTX_BLOCK_NAN (per-block + modulation capture), LTX_DIT_STOP_AT_BLOCK / LTX_BLOCK_STOP_AT_SUBOP truncation bisect, LTX_CHAIN_SAVE_DIR latent banking for replay-from-failure.
This was referenced Jul 6, 2026
Open
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
3 participants
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.
.cacheto.gitignore