docs(readme): News carries MXFP4 vLLM parity and the Vulkan e2e milestone - #160
Merged
Conversation
…tone The News section had gone stale: it still led with the audio-input entry and carried nothing about the two things that actually landed since. Both claims below are grounded in committed records, not restated ambition. - MXFP4: Qwen3-8B MXFP4 runs W4A16 Marlin by default (no env var needed), the greedy output matches the vLLM oracle, and steady-state decode is 45.45 tok/s against vLLM's 41.94 on GB10 — both graphed, both Marlin W4A16. - Vulkan: opt-125m greedy is STRICT token-exact against the oracle with 24 ops running native. Decode moved 8.59 -> 91.7 tok/s (10.7x). Stated as the correctness milestone it is: llama.cpp's Vulkan backend is still 2.62x ahead, and there is no Vulkan GPU on the lab box, so this is llvmpipe, not silicon. The oldest entry (audio input) retires to keep the landing page inside its 30000-char budget; the full history stays in docs/STATUS.md. FOLLOWING_AGENTS_PROTOCOL Assisted-by: Claude Opus 5 (1M context)
bakon11
pushed a commit
to bakon11/vllm.cpp
that referenced
this pull request
Aug 8, 2026
Two changes landed independently and each fit on its own: `a9a8581d` added a one-line MANIFESTO.md link to the header, and `mudler#160` rewrote the News block. Together they put README.md at 30056 chars, over the 30000-char budget, so check-readme-structure is red on main and every open PR inherits the failure. Retires the MiniMax-H3 News entry, the oldest of the three, leaving the two newest headlines (MXFP4 parity, Vulkan e2e) and 128 chars of headroom. The MiniMax-H3 result is unchanged and still recorded in docs/STATUS.md; News is a rolling window, not an archive. Also records MANIFESTO.md in .agents/mission.md, which is the landing-page positioning source. `a9a8581d` pointed the landing page at a new positioning document without recording it here, which is the half of that change the POL-DOC checkpoint asks for; this supplies it. I caused half of this by running check-doc-checkpoint against the working tree instead of the committed HEAD, which is exactly the failure mode our own notes warn about: that checker reads HEAD, so a pre-commit run cannot see the commit it is meant to judge. FOLLOWING_AGENTS_PROTOCOL Assisted-by: Claude Opus 5 (1M context)
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
2 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.
The README
## Newssection had gone stale — it still led with the audio-input entry and said nothing about the two things that actually landed since.Both new entries are grounded in committed records, not restated ambition:
VT_*vars are=0rollback switches only). Greedy output matches the vLLM oracle, and steady-state decode is 45.45 tok/s vs vLLM 41.94 on GB10, both graphed, both Marlin W4A16.opt-125mgreedy is STRICT token-exact vs the oracle with 24 ops native. Decode moved 8.59 → 91.7 tok/s (10.7×). Stated as the correctness milestone it is: llama.cpp Vulkan is still 2.62× ahead, and the lab box has no Vulkan GPU, so this is llvmpipe, not silicon.The oldest entry (audio input) retires to keep the landing page inside its 30000-char budget (now 29998); the full history stays in
docs/STATUS.md.Full doc-gate battery green: readme-structure, public-doc-tables, doc-checkpoint, now-current, state-order, agent-record, protocol-consistency, env-doc, policy.