docs: Q2_0 CUDA merged upstream; dspark drafter still fork-only

[epoyraz]

The ternary migration table still lists CUDA as 🔄 In review upstream, but ggml-org/llama.cpp#25707 ("CUDA: add Q2_0 support") was merged into mainline llama.cpp on 2026-07-30 (merge commit 9b2a088).

1. Mark CUDA merged

• Backend table — CUDA row: 🔄 In review upstream → ✅ Merged in mainline llama.cpp. Same PR link, unchanged.
• Summary paragraph below the table — it said "CUDA is the last one in review upstream; until it merges, use this demo". Now names all four merged backends and points at x86 AVX-512-VNNI as the only outstanding item (x86 already runs via the generic scalar CPU path from #24448). The demo's role is preserved but restated on its real terms: it still ships the fork binaries and the group-128 *-Q2_0.gguf files, which mainline cannot load.
• Two stale backend enumerations in the same section — the *-Q2_0_g64.gguf "Runs on" cell ("CPU and Metal so far") and the group-64 download intro ("CPU or Metal"). Both predate the Vulkan merge (#25430) and would have read as contradicting the corrected table; both now say CPU, Metal, Vulkan, CUDA.
• Intro line (§ Bonsai 27B) — dropped "with CUDA in review".

2. Say that the dspark drafter is still fork-only

The first commit alone would have made things worse: "CPU, Metal, Vulkan, and CUDA now run Q2_0 on mainline, no fork needed" reads as if the fork is now optional everywhere. It is not — speculative decoding still requires it, and neither README nor SPECULATIVE.md said so anywhere.

This is not hypothetical. On the upstream thread, @MaxKruse tried the drafter against the merged CUDA backend and hit:

gguf_init_from_reader: tensor 'dspark.fc.weight' has offset 337718592, expected 357584192
llama_model_load: error loading model: failed to load model from Ternary-Bonsai-27B-dspark-Q4_1.gguf
srv  load_model: failed to load draft model

Per @khosravipasha's reply on that thread, the dspark implementation is experimental, not ready to be upstreamed, and waiting on llama.cpp's own dspark support to merge and stabilize. So the second commit adds:

• A paragraph after the "no fork needed" line scoping it to plain Q2_0 inference and stating that speculative decoding is fork-only, with the error string included so it is greppable.
• The same note at the top of SPECULATIVE.md, which previously never mentioned mainline at all.
• A one-line caveat on the README's speculative-decoding bullet.

Scope

Docs only — 2 files, no scripts, binaries, or release tags touched. The pinned fork release prism-b9596-9fcaed7 and the group-128 files setup.sh downloads are unaffected.

Two things left deliberately alone, since you mentioned a major fork refactor is planned once #25707 lands:

• The *-PQ2_0.gguf row still says "Not supported yet / fork support planned" — that is presumably what the refactor decides.
• No claim is made about when dspark reaches mainline, only that it is blocked on upstream's own implementation.

Happy to trim this to the one-line table fix if you would rather keep the PR minimal.

🤖 Generated with Claude Code

[khosravipasha]

Thanks for the changes, will review and merge some portion of it. For the migration have not decided what's the most clean way yet.

[Copilot]

Pull request overview

Updates documentation for upstream CUDA Q2_0 support and the bundled dspark drafter’s fork dependency.

Changes:

• Marks CUDA Q2_0 support as merged upstream.
• Expands supported backend listings.
• Documents that the bundled dspark GGUF remains fork-only.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
README.md	Updates backend status and speculative-decoding caveats.
SPECULATIVE.md	Adds the bundled drafter’s mainline compatibility warning.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[khosravipasha]

Made some changes on our side and will merge.