…ry results

…sors are the sames, will be usefull to compare inputs and outputs

…rs for stage 1 and 2

…2 audio_to_video_attn video_to_audio_attn. Some final re-run to be done to double check

…ideo_to_audio_attn (no-LoRA)

…to-end block. Several steps are described in block0_reverse_engineering_map.md, Step M1 is done.

…eat for LoRA 0.5

…icAVTransformerBlock

…g the right inputs like the production python path

…umerical drift due to XLA op fusions. Distribution of error consistent with a drift

… or double free. As of now, forces exe synchronous calls.

…steps)

…ed early for positions, masks and clean latents

…ve registry and store

…a_noise_buf, v_latent_buf and a_latent_buf

…r blocks

…nd early deinit of preprocessing buffers

Use Flash Attention whenever this is possible. As of now it is wired for :
* self attention (audio and video)
* cross attention (audio to video and video to audio)

Not yet implemented for text cross attention (investigation needed because of the use of masks)

… H100 (#515)

Route video and audio text cross-attention (`attn2`/`audio_attn2`)
through FlashAttention3 when available, matching the existing FA3 path
used by self-attention and AV cross-attention.

* Pass `attn_meta`/`attn_params` instead of mask for text cross-attn
calls
* Remove unused `mask` field from `ForwardOpts` and
`v_text_ctx_mask`/`a_text_ctx_mask` from `SharedInputs`
* Simplify `forwardImpl` dispatch (no dead mask guard)

Stage 1 execution: 81.2s → 75.6s (-6.9%), Stage 2 execution: 11.6s →
11.1s (-4.3%), total: 231.6s → 206.3s (-10.9%), total compile: 96.6s →
78.0s (-19.3%).

# Conflicts:
#	zml/tensor.zig

Remove f32 upcast around GELU in FeedForward: keep activation in native
bf16.

Eliminates two unnecessary dtype conversions per FF call (92 blocks × 2
FF calls per block × 30 steps for stage 1).

~0.7s execution saving.

Changes:

* Switch transformer block calls to async execution in Stage 1 and Stage
2 denoising loops
* Edit timing reporting summary table to the old `other` column is
actually split into {Compile, Load, Execute}
* Gate all GPU profiling behind `--profile` CLI flag
* Add `analyze_trace.py`: analyzes XLA trace JSON files with GPU kernel
categorization, de-overlap correction, and `--fix` flag to emit
corrected traces (so even overlapping GPU events appear on Perfetto)

Impact: ~1.9s saved per generation (GPU utilization 92.5% --> 99.2% for
Stage 1, and 98.4% --> 99.7% Stage 2)

Improve inline doc comments in `inference.zig` and `model.zig` 

Rewrite `README.md` and `ARCHITECTURE.md`