[merceod]

Stacked on #690 (pr-09-dit).

The second docs PR for the ROCm work. It documents the media (#688), compile/env (#689) and DiT (#690) PRs and edits sections the first docs PR (#687) adds, so it stacks on #690. Docs only.

• architecture.rst gets the MoE, media decode and encoder attention rows in the per-platform kernel table, and the ROCm inductor override as it ended up (ATEN GEMMs for captured buckets of 32 tokens or fewer, which only matters under -O2 since Reduce cold start time with broad compile tuning levels #817 made -O1 the default).
• installation.rst gets the packages ROCm needs for media (PyAV and soundfile in place of torchcodec), the compile-cache recipe, and conv search on a cold node. A saved MIOpen database is the recommendation, and MIOPEN_FIND_MODE=FAST is only described as an option for bf16 VAE deployments (numbers below). The post-install check now quotes Kokoro's real cold start.
• environment_variables.rst lists MIOPEN_FIND_MODE with the other ROCm library variables.
• serving.rst covers --gpus on AMD GPUs.
• models.rst fills in the "Running on AMD GPUs" status table from our correctness runs and the DiT runs.

NVIDIA: docs only.

Tests

What	Result
ruff check .	clean
sphinx-build -b html docs with autosummary_generate=0 and api.rst excluded (the full API build didn't finish in 40 minutes on our machine)	narrative pages build clean, no warnings
Rebase onto main 54f29da (2026-10-26, the whole stack)	CPU on a devel node, GPUs hidden: test/modular only fails the two test_dump_stacks environment cases main fails too (7174 passed before the last three test fixes, targeted rerun on the tip 398 passed), other CPU suites 248 passed, TP lockstep sim 7 passed, test/rust 368 passed. One MI300X: test/rocm 675 passed, fused-MoE and sampler GPU tests 228 passed, served smoke rows for Qwen3.5 0.8B, Kokoro, Whisper large-v3 and Cosmos3 DROID match the earlier runs

Numbers quoted in the pages

Conv search on a cold node. Qwen3-Omni (qwen3omni_colocated) on one MI325X, time to a healthy server: 1146 s with an empty MIOpen database, of which Code2Wav's captures take 645 s, and 223 s from a saved database (Code2Wav 101 s, Talker 44 s, Thinker 36 s). With MIOPEN_FIND_MODE=FAST and an empty database it took 1668 s. The vocoder asks for the full search (cudnn.benchmark), which FAST doesn't skip, and FAST's fallback kernels are 35-48% slower per vocoder step. FAST does cut a bf16 VAE's first decode on MI300X (Wan2.2 480x832x33 161 -> 6.6 s, Cosmos3 720p 82 -> 0.1 s, FLUX.2 autoencoder 83 -> 1.5 s, later bf16 decodes within 4%) but slows fp32 convs (Wan2.2's fp32 decode 3.6 -> 4.5 s).

Compile caches. Seeded node-local Inductor and MIOpen caches vs empty ones, time to a healthy server on one MI300X VM slice: Orpheus 1086 -> 137 s, Qwen3.5-9B 582 -> 156 s, Kokoro 1272 -> 171 s.

Not in this PR

The status table was caught up in the last docs commit here: the multi-GPU configs (fused all-reduce at TP2 and TP4 included), the models that had no row yet (bagel, nemotron_duplex, pi05, vjepa2_ac, flux2_klein_9b, cosmos3_super TP2, the Qwen3-TTS Base and VoiceDesign configs, Waypoint 720p) and the wan22 VAE note. The Qwen3-Omni text divergence it mentions didn't reproduce in 144 passes. pi05's SigLIP weight-loading fix merged as #696 and V-JEPA2's preprocessing fix is #697, neither ROCm-specific. The exclusive-node tables land after this.

Since r8

Since the r8 cut: Qwen3.5's AMD status row says prefix caching is off on ROCm (#686 keeps the recurrent prefix cache shut over AITER's GDN kernels, which keep no checkpoints).

Checks on this rebase

The stack now sits on main f7cb5ca (2026-10-29), every branch cut from one integration tree.

Check	Result
CPU on a devel node with the GPUs hidden: ruff, test/ci, the TP lockstep sim, test/rust, CPU Core	ruff clean, sim 14 passed, test/rust 436 passed, CPU Core fails nothing main doesn't (main's own 34: test_dump_stacks x2, loop_output_ownership x32)
test/rocm on one MI325X (a VM slice, torch 2.12.1+rocm7.2)	715 passed, 9 skipped, 0 failed
MoE and sampler tests there	MoE EP 53 passed, sampler 145 passed. The 75 test_*moe* failures are main's GLM-5.2 kernel tests, the same set on the previous head
Served rows there through the correctness runner	all PASS: Qwen3.5 0.8B and 9B text (no divergence away from near ties, 4/4 needles), Whisper large-v3 (WER 1.64%, 99/100 transcripts equal to the reference), Kokoro (WER delta 0), Qwen3-Omni colocated (50/50, spoken WER 0), pi0.5 and Cosmos3 DROID (two instances bit-same)
Qwen3.5 9B TP2 on two GPUs there	PASS (21/32 exact, 11 near ties, 0 unexplained), the fused all-reduce on both ranks

Nothing fails on this tree that the previous head doesn't: the failing-set diff against it is empty for test/rocm, MoE, the sampler and the rows.