[DO NOT MERGE — wrong repo] KakeyaLattice dissemination kit staged in AgentMemorySystem by mistake

.gitignore

@@ -0,0 +1,4 @@
+dissemination/kakeyalattice/huggingface/space/__pycache__/
+__pycache__/
+*.pyc
+dissemination/kakeyalattice/arxiv/arxiv_submission.tar.gz

dissemination/kakeyalattice/DISSEMINATION_PLAN.md

@@ -0,0 +1,88 @@
+# KakeyaLattice Dissemination Kit
+
+**Target repo**: [`FluffyAIcode/LLM-KV--Cache-compress`](https://github.com/FluffyAIcode/LLM-KV--Cache-compress)
+**Goal**: move the project from "搜不到" (discoverable only by exact name) to "natural-language discoverable" on the four primary channels researchers / engineers actually use.
+
+## Why this kit exists
+
+KakeyaLattice v1.4/v1.5 is fully measured and release-ready, but as of 2026-04-27 generic queries like *"lattice KV cache compression"*, *"E8 KV quant vLLM"*, or *"Kakeya-Zamir LLM"* return NexusQuant / NestQuant / KV-Compress / LMCache — not this repo. The four causes we can actually fix from the author side:
+
+1. GitHub repo has **zero topics** → excluded from `/topics/*` discovery pages.
+2. No arXiv ID → no Google Scholar / Semantic Scholar / Connected Papers index → no academic backlinks.
+3. No vLLM-ecosystem issue → not cross-referenced from the 76k-star vLLM repo (NexusQuant got this via `vllm#39241` and it's already its #1 inbound source).
+4. No HuggingFace Space and no Papers with Code entry → no `paperswithcode.com/paper/...` landing page and no HF hub search hit.
+
+This kit completes **what can be automated** (config files, LaTeX tarball builder, issue Markdown, Space scaffold, PwC JSON) and stages **what requires a human account** (arXiv endorsement + upload, HF CLI login + push, PwC submit button) as one-command steps.
+
+## Execution order (5 steps, ~30–40 min of human time total)
+
+| # | Task | Where it lives | Who executes | Time |
+|---|------|----------------|--------------|------|
+| 1 | Set GitHub topics + description | `github_topics/apply.sh` | repo owner, 1 command | 30 s |
+| 2 | Submit arXiv preprint | `arxiv/` | Allen Li, arXiv account | 10 min (+ endorsement wait) |
+| 3 | Open vLLM Discussion / Issue | `vllm_issue/BODY.md` | anyone with GitHub account | 2 min |
+| 4 | Deploy HuggingFace Space demo | `huggingface/space/` | any HF account | 5 min |
+| 5 | Submit Papers with Code entry | `paperswithcode/` | any PwC account | 3 min |
+
+After all five land, you should have **4 new inbound backlinks** (vLLM issue, HF Space, arXiv abstract page, PwC paper page) and **7 GitHub topic pages** pointing at the repo. Empirically this is the minimum needed to show up on natural-language LLM + KV-cache queries.
+
+## Per-step quick start
+
+```bash
+# 1. GitHub topics (run from any machine with gh CLI auth'd as repo owner)
+bash dissemination/kakeyalattice/github_topics/apply.sh
+
+# 2. Build arXiv tarball (produces arxiv_submission.tar.gz, upload at arxiv.org/submit)
+bash dissemination/kakeyalattice/arxiv/build_tarball.sh
+# Then follow dissemination/kakeyalattice/arxiv/SUBMIT.md
+
+# 3. Open vLLM issue (body ready at vllm_issue/BODY.md)
+gh issue create -R vllm-project/vllm \
+    --title "$(cat dissemination/kakeyalattice/vllm_issue/TITLE.txt)" \
+    --body-file dissemination/kakeyalattice/vllm_issue/BODY.md
+
+# 4. Deploy HF Space
+bash dissemination/kakeyalattice/huggingface/deploy.sh   # requires `huggingface-cli login`
+
+# 5. Submit to Papers with Code (manual, 30 seconds) — see paperswithcode/SUBMIT.md
+```
+
+## Files in this kit
+
+```
+dissemination/kakeyalattice/
+├── DISSEMINATION_PLAN.md        ← this file
+├── github_topics/
+│   ├── topics.json              ← topic list (source of truth)
+│   ├── description.txt          ← GitHub "About" one-liner
+│   └── apply.sh                 ← `gh` CLI command, one-shot
+├── arxiv/
+│   ├── SUBMIT.md                ← submission walkthrough (endorsement, categories)
+│   ├── metadata.yaml            ← title, authors, abstract, categories, comment
+│   ├── build_tarball.sh         ← produces arxiv_submission.tar.gz from reports/paper/
+│   └── endorsement_request.md   ← template email to request cs.LG endorsement
+├── vllm_issue/
+│   ├── TITLE.txt                ← issue title
+│   ├── BODY.md                  ← issue body (mirrors NexusQuant vllm#39241 format)
+│   └── LABELS.txt               ← recommended labels
+├── huggingface/
+│   ├── space/                   ← full HF Space repo scaffold (app.py, requirements.txt, README.md)
+│   ├── deploy.sh                ← pushes Space to hf.co/spaces/<user>/kakeyalattice
+│   └── MODEL_CARD_EDIT.md       ← snippet to add to any HF model card that benefits from KakeyaLattice
+└── paperswithcode/
+    ├── SUBMIT.md                ← submit walkthrough
+    ├── entry.json               ← paper + code + results (copy-paste ready)
+    └── sota_tables.md           ← pre-filled iso-PPL and iso-bit leaderboard rows
+```
+
+## Measurement of success
+
+After execution, re-run these natural-language queries; each should surface the repo or its arXiv page in the first result page (currently zero do):
+
+- `lattice KV cache compression vLLM`
+- `E8 lattice KV cache quantization`
+- `Kakeya-Zamir nested lattice LLM`
+- `D4 E8 KV cache H200`
+- `KV cache compression plugin vLLM 2026`
+
+We expect first Google indexing of the arXiv page within **24–72 h** and first Bing/DuckDuckGo within **5–7 days** post-submission. GitHub topics update is immediate. HF Space and PwC typically index within 24 h.

dissemination/kakeyalattice/README_PATCH.md

@@ -0,0 +1,98 @@
+# README patch for FluffyAIcode/LLM-KV--Cache-compress
+
+Paste this block directly below the first heading (`# KakeyaLattice — v1.4
+KV-Cache Compression`) in the KakeyaLattice repo's `README.md`. Every
+badge is self-updating: they reflect live status as soon as the
+corresponding step in the dissemination kit is completed.
+
+```markdown
+[![Release v1.5](https://img.shields.io/github/v/release/FluffyAIcode/LLM-KV--Cache-compress?color=blue&label=release)](https://github.com/FluffyAIcode/LLM-KV--Cache-compress/releases/latest)
+[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-green.svg)](LICENSE)
+[![arXiv](https://img.shields.io/badge/arXiv-pending-b31b1b.svg)](reports/paper/kakeyalattice.pdf)
+[![Papers with Code](https://img.shields.io/badge/Papers_with_Code-pending-21cbce.svg)](https://paperswithcode.com/paper/kakeyalattice)
+[![HF Space](https://img.shields.io/badge/%F0%9F%A4%97-demo-yellow.svg)](https://huggingface.co/spaces/FluffyAIcode/kakeyalattice)
+[![vLLM Issue](https://img.shields.io/badge/vLLM-feature_request-informational.svg)](https://github.com/vllm-project/vllm/issues?q=KakeyaLattice)
+
+**Topics**: `kv-cache` · `kv-cache-compression` · `quantization` · `vllm` ·
+`lattice-quantization` · `e8-lattice` · `d4-lattice` · `nested-lattice` ·
+`llm-inference` · `long-context` · `h200`
+```
+
+After arXiv lands, replace the `arXiv-pending` badge line with:
+
+```markdown
+[![arXiv](https://img.shields.io/badge/arXiv-26MM.NNNNN-b31b1b.svg)](https://arxiv.org/abs/26MM.NNNNN)
+```
+
+and add a **Citation** section at the bottom of `README.md`:
+
+```markdown
+## Citation
+
+If you use KakeyaLattice in your research, please cite:
+
+​```bibtex
+@misc{li2026kakeyalattice,
+  author       = {Allen Li},
+  title        = {{KakeyaLattice}: Nested-Lattice {KV}-Cache Compression
+                  with {K}akeya-Style Discrete Codebooks ({D}4 + {E}8 Joint Release)},
+  year         = {2026},
+  eprint       = {26MM.NNNNN},
+  archivePrefix= {arXiv},
+  primaryClass = {cs.LG},
+  url          = {https://arxiv.org/abs/26MM.NNNNN},
+  note         = {Code: \url{https://github.com/FluffyAIcode/LLM-KV--Cache-compress}}
+}
+​```
+```
+
+## One-command re-dissemination
+
+Add this section somewhere near the end of `README.md`:
+
+```markdown
+## Dissemination
+
+To keep the project discoverable (GitHub topics, arXiv, vLLM issue, HF
+Space, Papers with Code), use the dissemination kit shipped in
+[`dissemination/`](dissemination/DISSEMINATION_PLAN.md).  All five
+channels are scripted to one command each:
+
+​```bash
+# 1. GitHub topics + description (requires repo-admin gh CLI auth)
+bash dissemination/github_topics/apply.sh
+
+# 2. arXiv submission tarball (upload at https://arxiv.org/submit)
+bash dissemination/arxiv/build_tarball.sh
+
+# 3. Open a vLLM issue (body pre-written)
+gh issue create -R vllm-project/vllm \
+    --title "$(cat dissemination/vllm_issue/TITLE.txt)" \
+    --body-file dissemination/vllm_issue/BODY.md
+
+# 4. Deploy HF Space (requires huggingface-cli login)
+bash dissemination/huggingface/deploy.sh
+
+# 5. Submit to Papers with Code (manual form, 3 min)
+#    entries ready at dissemination/paperswithcode/entry.json
+```
+
+## Where to drop the kit
+
+The kit currently lives in the `AgentMemorySystem` repo (branch
+`AgentMemory/kakeyalattice-dissemination-kit-f31f`). To adopt it into
+KakeyaLattice:
+
+```bash
+cd LLM-KV--Cache-compress
+git remote add ams https://github.com/FluffyAIcode/AgentMemorySystem
+git fetch ams AgentMemory/kakeyalattice-dissemination-kit-f31f
+git checkout ams/AgentMemory/kakeyalattice-dissemination-kit-f31f -- \
+    dissemination/kakeyalattice
+git mv dissemination/kakeyalattice dissemination
+git commit -m "Adopt KakeyaLattice dissemination kit"
+git push
+```
+
+From then on, all five steps are re-runnable from inside the KakeyaLattice
+repo with no re-staging.

dissemination/kakeyalattice/arxiv/SUBMIT.md

@@ -0,0 +1,103 @@
+# arXiv submission walkthrough — KakeyaLattice
+
+Est. time: **10 minutes** of active work + endorsement wait (hours to days
+if first-time cs.LG submitter, instant if already endorsed).
+
+## Prerequisites
+
+- An arXiv account (register at https://arxiv.org/user/register)
+- cs.LG endorsement (if first-time — see `endorsement_request.md`)
+- LaTeX toolchain (`pdflatex`, `bibtex`) — optional but recommended
+
+## Step 1 — Build the submission tarball
+
+From the KakeyaLattice repo root:
+
+```bash
+bash dissemination/kakeyalattice/arxiv/build_tarball.sh
+```
+
+Output: `dissemination/kakeyalattice/arxiv/arxiv_submission.tar.gz`
+
+Sanity-check:
+
+```bash
+tar -tzf dissemination/kakeyalattice/arxiv/arxiv_submission.tar.gz | head -20
+```
+
+You should see `kakeyalattice.tex` and (if pdflatex was available) a
+pre-built `kakeyalattice.bbl`.
+
+## Step 2 — Fill the submission form
+
+Go to https://arxiv.org/submit → "Start a new submission".
+
+Paste values from `metadata.yaml`:
+
+| Form field | Value source |
+|---|---|
+| Title | `title` |
+| Author(s) | `authors` (single author: Allen Li) |
+| Abstract | `abstract` (paste as-is; arXiv strips LaTeX automatically) |
+| Comments | `comments` |
+| Primary subject | **cs.LG** |
+| Cross-listing | cs.CL, cs.IT, cs.DS |
+| MSC class | 94A29, 68T50 |
+| ACM class | I.2.7; E.4 |
+| License | **CC BY 4.0** (recommended) |
+
+## Step 3 — Upload tarball
+
+- Choose "Upload: tar archive of sources"
+- Upload `arxiv_submission.tar.gz`
+- Wait for server-side build (typical: 2–5 min)
+- If build fails: the error log usually points to a missing figure or package;
+  copy it into the tarball and rebuild.
+
+## Step 4 — Preview PDF
+
+arXiv auto-generates a preview PDF. Compare against the source PDF at
+`reports/paper/kakeyalattice.pdf`; they should be visually identical. If the
+preview is missing references or figures, fix the tarball and resubmit.
+
+## Step 5 — Submit
+
+Click "Submit" on the metadata page. You'll get an immediate confirmation
+with a temporary ID (like `submit/12345678`). The permanent
+`arXiv:26MM.NNNNN` ID is assigned after the next daily announcement cycle
+(Monday–Thursday announce at 20:00 UTC; Friday's submissions announce Monday).
+
+## Step 6 — After publication
+
+Once you have the arXiv ID, update KakeyaLattice in this order:
+
+```bash
+# In FluffyAIcode/LLM-KV--Cache-compress:
+
+# 6a. Badge + citation in README
+# Add to top of README.md:
+#   [![arXiv](https://img.shields.io/badge/arXiv-26MM.NNNNN-b31b1b.svg)](https://arxiv.org/abs/26MM.NNNNN)
+
+# 6b. Update Papers with Code entry (see ../paperswithcode/)
+# 6c. Update HF Space README badge (see ../huggingface/space/README.md)
+# 6d. Post the arXiv link as a comment on the vLLM issue (see ../vllm_issue/)
+# 6e. Reply to NestQuant / NexusQuant threads with the arXiv link for reverse backlinks
+```
+
+Google Scholar usually indexes within **24–48 h** of arXiv publication.
+Semantic Scholar and Connected Papers within **1–3 days**.
+
+## Common pitfalls
+
+- **Non-ASCII characters** in the abstract field: replace em-dashes (—) with
+  double-hyphens (--), and curly quotes with straight quotes. metadata.yaml
+  already does this.
+- **Missing `.bbl`**: if arXiv can't find your bibliography, either
+  pre-build it (the script does this when pdflatex is available) or include
+  the `.bib` file and ensure `\bibliography{kakeyalattice}` points to it.
+- **Figures > 6 MB**: compress PDFs with `gs -sDEVICE=pdfwrite
+  -dPDFSETTINGS=/ebook`.
+- **Version update**: if you revise the paper post-publication (v1.5 adds
+  new data, for example), submit as a **replacement** from the same abstract
+  page, not as a new submission. Each version gets `v1`, `v2` suffixes under
+  the same arXiv ID.

dissemination/kakeyalattice/arxiv/build_tarball.sh

@@ -0,0 +1,72 @@
+#!/usr/bin/env bash
+# Build an arXiv-compliant tarball from reports/paper/.
+#
+# Usage (run from the KakeyaLattice repo root):
+#   bash dissemination/kakeyalattice/arxiv/build_tarball.sh
+# Produces: dissemination/kakeyalattice/arxiv/arxiv_submission.tar.gz
+#
+# The tarball contains:
+#   - kakeyalattice.tex          (main source)
+#   - any .bbl / .bib / figures / style files from reports/paper/
+# and omits build artefacts listed in reports/paper/.gitignore.
+#
+# Requirements: bash, tar, grep, awk; pdflatex+bibtex only if you want
+# to pre-build the .bbl (recommended, arXiv builds faster with .bbl included).
+
+set -euo pipefail
+
+HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "$HERE/../../.." && pwd)"   # dissemination/kakeyalattice/arxiv/ -> repo root
+PAPER_DIR="$REPO_ROOT/reports/paper"
+OUT="$HERE/arxiv_submission.tar.gz"
+STAGE="$(mktemp -d)"
+
+if [[ ! -f "$PAPER_DIR/kakeyalattice.tex" ]]; then
+    echo "ERROR: expected $PAPER_DIR/kakeyalattice.tex" >&2
+    echo "Run this script from inside the KakeyaLattice repo (FluffyAIcode/LLM-KV--Cache-compress)." >&2
+    exit 1
+fi
+
+echo "==> Staging paper sources in $STAGE"
+cp "$PAPER_DIR"/*.tex "$STAGE/"
+cp "$PAPER_DIR"/*.bib "$STAGE/" 2>/dev/null || true
+cp "$PAPER_DIR"/*.cls "$STAGE/" 2>/dev/null || true
+cp "$PAPER_DIR"/*.sty "$STAGE/" 2>/dev/null || true
+
+# Figures subdirs (common layouts)
+for d in figures figs images img; do
+    if [[ -d "$PAPER_DIR/$d" ]]; then
+        cp -r "$PAPER_DIR/$d" "$STAGE/"
+    fi
+done
+
+# Try to pre-build the .bbl so arXiv's build path is shorter.
+if command -v pdflatex >/dev/null && command -v bibtex >/dev/null; then
+    echo "==> Pre-building .bbl with pdflatex+bibtex"
+    pushd "$STAGE" >/dev/null
+    pdflatex -interaction=nonstopmode kakeyalattice.tex >/dev/null || true
+    bibtex kakeyalattice >/dev/null || true
+    pdflatex -interaction=nonstopmode kakeyalattice.tex >/dev/null || true
+    pdflatex -interaction=nonstopmode kakeyalattice.tex >/dev/null || true
+    # Remove intermediate artefacts; keep .bbl
+    rm -f *.aux *.log *.out *.toc *.fls *.fdb_latexmk *.synctex.gz *.blg
+    popd >/dev/null
+else
+    echo "WARN: pdflatex/bibtex not found — arXiv will build the .bbl server-side."
+fi
+
+echo "==> Creating tarball $OUT"
+rm -f "$OUT"
+tar -czf "$OUT" -C "$STAGE" .
+ls -lh "$OUT"
+
+echo
+echo "Next steps:"
+echo "  1. Go to https://arxiv.org/submit and start a new submission"
+echo "  2. Primary category: cs.LG (see metadata.yaml)"
+echo "  3. Upload $OUT as 'tar archive of sources'"
+echo "  4. Paste title / abstract / comments from metadata.yaml"
+echo "  5. License: CC BY 4.0 (recommended)"
+echo
+echo "If this is your first cs.LG submission, request endorsement first:"
+echo "  see dissemination/kakeyalattice/arxiv/endorsement_request.md"