llama-modchip v2.1 & 256K CPU Co-Design: Running 30B Hybrid SSM-MoE & GQA Models on Pure CPU at 20–60+ tok/s (0 GPUs)

llama-modchip ("Priiloader for CPU GGUF Inference") + RouteCrack & Stockfish-PVS v2 (M=7): How a 1.64 MB Chess-Engine Search Inside L3 Cache + 2MB HugePages + _MM_HINT_NTA Unlocks 48.4–60.4 tok/s on Nemotron-3.5-30B-A3B (256K Context) and +38.5% Decode on Qwen3-30B-A3B with Zero GPUs

📄 Documentation Navigation:

• Overview, 60-Second Quickstart, Stockfish-PVS vs Neural Draft & Leaderboard (README.md) — You are here
• Full Technical Paper & Mathematical Proofs (ARCHITECTURE.md) — Deep dive into Mamba-2/GQA physics, _MM_HINT_NTA, 2MB HugePages STLB math, LLAMA_MOE_TOP_K=4 norm preservation, RouteCrack + Stockfish-PVS v2 break-even theorem, negative ablations, and Linux Kernel auditing benchmarks
• RouteCrack + Stockfish-PVS v2 Catalog of 96+ Techniques (speculative/ROUTECRACK_TECHNIQUES.md) — How Chess-Engine PVS + Zobrist Transposition Tables unlocked 4 dormant RouteCrack techniques (M=7 Singular Horizon Extension, 60.40 tok/s peak, mean len = 6.90)
• Engineering FAQ & Xeon/EPYC Server Scaling (FAQ.md) — SWA vs Full 256K NIAH retrieval, Mamba-2 zero-rollback gating, Stockfish-PVS L3 cache physics, and Intel Xeon Scalable / AMD EPYC multi-agent scaling
• Release Notes (RELEASE_NOTES.md) — What's new in v2.1.0 (Stockfish-PVS v2 Sweet Spot, Apple Silicon M4 Pro & Fedora Ryzen 8745HS support)

1. What This Repo Does (Direct Summary & Live Production Numbers)

• The Goal: Run 30B-class Sparse MoE LLMs on a pure CPU (0 GPUs) at 20–60+ tok/s with up to a 262,144-token (256K) context window for real autonomous coding (Claude Code with 40 tools).
• How It Works: 9 low-level C++ / Linux kernel optimizations for llama.cpp (ggml-cpu), paired with RouteCrack + Stockfish-PVS v2 (Sweet Spot, M=7) (a draftless C++ speculative engine combining an 8-level FNV-64 hash cascade, Online Trie, Macro Codebook, and a 1.64 MB L2/L3-resident Chess-Engine Principal Variation Search / Zobrist Transposition Table executing in < 0.005 ms) and llama-modchip (v0.9 Public Beta) — a "Priiloader for CPU GGUF Inference" (scripts/llama_modchip.py) that inspects any .gguf header + CPU topology in < 5 ms, engages architecture Safety Locks, and tunes llama-server live via /dev/shm/llama_modchip.bin.
• What Is Tested vs. Beta:
  • Certified in Daily Production (Tier 1): AMD Ryzen 9 9955HX (Zen 5, 16C/32T, Dual-CCD 64 MB L3, AVX-512 VNNI) + 64 GB DDR5 + Linux 6.8+ (2MB THP), tested across 4 real GGUF architectures (Nemotron-3.5-30B-A3B at 256K context, Qwen3-30B-A3B at +38.5% base decode, Nemotron-H-8B, and Llama-40L).
  • Public Beta & Community Platforms (Tier 2): Auto-adaptive support for AMD Zen 4 / Zen 3 / EPYC (including verified fedora/ setup on AMD Ryzen 7 8745HS GEEKOM A8), Intel AVX2 / AVX-512 CPUs, and apple-silicon/ (Apple M4 Pro 24 GB Metal at 47–49 tok/s up to 128K).

⚡ 60-Second Quickstart (./modchip Unified CLI, Live Web HUD & Terminal Chat)

# 1. Clone & run the automated installer (builds patched llama.cpp + Stockfish-PVS & downloads GGUF):
git clone https://github.com/synchopate/nemotron-30b-cpu-256k.git && cd nemotron-30b-cpu-256k
./modchip setup                  # Or Bring-Your-Own-GGUF: ./modchip setup --skip-model --model /path/to/any.gguf

# 2. Start the 256K CPU Engine + Live Web Telemetry HUD in the background:
./modchip start                  # Or pass any GGUF: ./modchip start /path/to/model.gguf

# 3. Pick how you want to interact & verify your tok/s immediately:
./modchip chat                   # 💬 Interactive terminal chat with live tok/s & Stockfish-PVS telemetry
./modchip bench                  # 📊 30s automated 2-stage benchmark + shareable CPU scorecard
./modchip web                    # 🌐 Open Live Web Chat & Telemetry HUD at http://127.0.0.1:8787
./modchip claude                 # 🤖 Launch Claude Code CLI wired to the local 256K CPU engine

♟️ Why a 1.64 MB Stockfish Chess Engine in L3 Cache Beats a 1B Neural Draft Model on CPU

On GPUs (3,350 GB/s HBM3), running a 1B neural draft model takes < 1 ms. On a CPU (71.7 GB/s Dual-Channel DDR5), small neural draft models fail catastrophically because streaming 600 MB of 1B draft weights for 5 steps (3.0 GB from DDR5) takes 40 ms — slower than running the 30B Sparse MoE target model itself (1.8 GB active $\rightarrow 25\text{ ms}$), while completely flushing the 64 MB L3 cache!

Instead, we paired RouteCrack (an 8-level FNV-64 Hash Cascade + Online Session Trie + 11,382-entry Macro Codebook acting as an $O(1)$ candidate move generator) with StockfishDrafter (rc_stockfish.h) — a Chess-Engine Principal Variation Search (PVS) + 4-Way Bucketed Zobrist Transposition Table that lives in 1.64 MB (2.5% of L3 cache):

🏆 Verified Community CPU & Unified-Memory Leaderboard

Run ./modchip bench on your machine and ➕ Click Here to Submit Your CPU Scorecard (1-Click GitHub Issue) to be added to the table below!

2. llama-modchip (v0.9 Public Beta): "Priiloader for CPU GGUF Inference" & Verified Hardware/Software Matrix

Inspired by console homebrew patchers (Priiloader / hacks.ini), scripts/llama_modchip.py turns our 9 low-level C++/kernel optimizations into a generalized, safety-locked auto-profiler and /dev/shm runtime controller that operates at two complementary levels across any GGUF model:

1. Startup & Graph-Construction Profile (< 5 ms GGUF + Topology Inspect): Parses the binary .gguf header (general.architecture, ssm.* vs attention.head_count_kv, expert_count, expert_used_count, expert_weights_norm) and Linux host topology (/proc/cpuinfo ISA flags, /sys/devices/system/cpu/cpu*/cache/index3 L3 CCD domains, physical core IDs, 2MB THP) to configure load-time graph parameters (LLAMA_MOE_TOP_K, --repack, MADV_HUGEPAGE, 1:1 Core Pinning, -ctk q8_0 -ctv q4_0).
2. Per-Forward-Pass Live /dev/shm/llama_modchip.bin Control: Exposes a 48-byte POSIX shared-memory control block read on every forward/draft pass by ggml-cpu/ops.cpp and common/ngram-map.cpp, enabling zero-restart runtime tuning of SWA window (attn_window), attention sinks (attn_sinks), interleaved stride (swa_stride), L3 _MM_HINT_NTA boundary (l3_window), and RouteCrack certainty gates (rc_enabled, rc_cert_d0/d1/d2).
3. Architecture Safety Locks: Automatically blocks optimizations that would corrupt a model's memory (e.g., blocking full 4K SWA on pure GQA Transformers and switching them to fail-closed 3:1 Interleaved SWA via LLAMA_SWA_STRIDE=4, or locking MoE Top-K pruning on dense models).

🔬 Public Beta Disclosure: Where It Is Tested & Certified

[!IMPORTANT]
llama-modchip v0.9 is currently in Public Beta. Below is the exact hardware, OS/kernel stack, and GGUF model matrix verified in production and live benchmarks, versus configurations where we are actively seeking community benchmarks.

📊 Live Multi-Model Verification Matrix (python3 scripts/llama_modchip.py --audit)

Run the 1-Command Hardware Check & llama-modchip Inspector:

# 1. Zero-dependency CPU readiness & DRAM bandwidth estimator (no GGUF required):
curl -sL https://raw.githubusercontent.com/synchopate/nemotron-30b-cpu-256k/main/scripts/check_cpu_readiness.py | python3

# 2. Inspect ANY GGUF model, apply live /dev/shm/llama_modchip.bin hooks & print optimal launch command:
python3 scripts/llama_modchip.py --model /path/to/model.gguf --apply-shm --print-cmd

# 3. Run 1-command A/B benchmark (Stock vs llama-modchip), multi-model audit, or interactive Curses TUI:
python3 scripts/llama_modchip.py --model /path/to/model.gguf --bench
python3 scripts/llama_modchip.py --audit model1.gguf model2.gguf
python3 scripts/llama_modchip.py --model /path/to/model.gguf --tui

3. The 9 Microarchitectural Optimizations (llama-modchip Stack)

Every single optimization is documented with full mathematical derivations and C/C++ code in ARCHITECTURE.md (§3.1–§3.10) and controlled at runtime by scripts/llama_modchip.py:

1. Architecture-Aware Sliding Window Attention + 16 Sinks (LLAMA_ATTN_WINDOW=0 vs 4096, LLAMA_SWA_STRIDE) (ARCHITECTURE.md §3.1 & §3.10)
   • Full 256K Exact Mode (LLAMA_ATTN_WINDOW=0, Default in systemd/nemotron-256k.service) vs. SWA Mode (LLAMA_ATTN_WINDOW=4096, LLAMA_SWA_STRIDE=1): Setting LLAMA_ATTN_WINDOW=0 retains exact 6-layer GQA attention across all 262,144 tokens (occupying 384 MB raw / 432 MB in symmetric Q4_0/Q4_0 at 1.73 KB/tok, ~432 MB in dynamic entropy-gated EAKV Q8_0/Q4_0, or 624 MB in static --cache-type-k q8_0 --cache-type-v q4_0 at 2.44 KB/tok, streamed via _MM_HINT_NTA). Setting LLAMA_ATTN_WINDOW=4096 bounds GQA attention to [0..15] U [cur_pos-4096..cur_pos] (6.31–9.76 MB hot working set inside L3), relying on the 46 non-attention blocks (23 Mamba-2 SSM + 23 Sparse MoE) for >4K global state.
   • On Pure GQA (Qwen3-30B-A3B, Llama): llama-modchip engages a Safety Lock (LLAMA_SWA_STRIDE=4) — applying L3 SWA to 3 out of 4 layers (75%) while keeping every 4th layer (25%) at full 256K global attention.
2. Bidirectional L3 Cache Isolation via _MM_HINT_NTA (LLAMA_KV_L3_WINDOW=2048) (ARCHITECTURE.md §3.2)
   Distinct from LLAMA_ATTN_WINDOW (which bounds the attention horizon), LLAMA_KV_L3_WINDOW=2048 controls the hardware cache-line prefetch hint inside ggml-cpu/ops.cpp: recent KV tokens within LLAMA_KV_L3_WINDOW are prefetched with temporal _MM_HINT_T0 (kept warm in L3), while older KV blocks and the 1.88 GB/token MoE weight stream in arch/x86/quants.c use non-temporal _MM_HINT_NTA (LLAMA_KV_L3_WINDOW auto-sized to $\le 25%$ of L3 capacity).
3. ISA-Adaptive In-Memory Weight Tile Repacking (--repack / q4_0_8x8) (ARCHITECTURE.md §3.3)
   Eliminates 30% bit-unpacking overhead via single-cycle VPDPBUSD (AVX-512 VNNI) or AVX2 matrix tiles (+19% GEMV/GEMM).
4. 2MB Transparent HugePages (MADV_HUGEPAGE Arena + MMap) (ARCHITECTURE.md §3.4)
   Collapses 4.9M 4KB pages into ~9,300 2MB pages (19.04 GiB), expanding per-core L2 STLB coverage from 12 MB to 6.14 GB and eliminating 24% page-table walk overhead across sparse MoE jumps.
5. Dual-CCD 1:1 Physical Core Pinning + Static GQA Partition (LLAMA_KV_DISABLE_CHUNKING=1) (ARCHITECTURE.md §3.5)
   Locks OpenMP threads 1:1 to non-SMT physical cores (OMP_PROC_BIND=close, OMP_PLACES=cores), eliminating SMT contention and 40ns cross-CCD Infinity Fabric L3 snoops (306.8 tok/s prefill).
6. Draftless C++ Speculative Engine (RouteCrack + Stockfish-PVS v2 via --spec-type ngram-simple --spec-ngram-simple-size-m 7) (ARCHITECTURE.md §3.6 · speculative/ROUTECRACK_TECHNIQUES.md)
   common/ngram-map.cpp overrides common_ngram_simple_draft() to hijack --spec-type ngram-simple and execute the L3-resident C++ RouteCrack + Stockfish-PVS v2 ensemble (speculative/cpp/rc_stockfish.h, speculative/cpp/rc_predictor.h, speculative/cpp/rc_trie_v2.h: 8-level FNV-64 Hash Cascade + Online Session Trie Rescue + 11,382-Entry Macro Codebook + Chess-Engine Principal Variation Search / Zobrist Transposition Table + Two-Zone 16K Hot / 24K Prefix Matcher, loaded from RC_TABLES_DIR) in < 0.005 ms inside 1.64 MB of L2/L3 cache. By cross-validating statistical predictions against the live-session Zobrist TT (tt5/tt4/tt3 + Countermove + Continuation History), Stockfish-PVS v2 unlocks 4 dormant RouteCrack techniques (including Singular Horizon Extension M=7), delivering +223 accepted tokens with 0 extra wasted tokens (-32 Mamba-2 rollbacks) on our 77,456-token held-out benchmark and reaching 48.39 tok/s average / 60.40 tok/s peak (mean len = 6.90 tokens/step, 96.4% acceptance) on live 25K-context code refactoring and 85.0% draft acceptance on technical prose!
7. Entropy-Guided Asymmetric KV Quantization (EAKV + VPI, -ctk q8_0 -ctv q4_0) (ARCHITECTURE.md §3.7)
   Cuts KV cache memory footprint and DRAM bandwidth by 58%–62.5% (384 MB raw / 432 MB in symmetric Q4_0, ~432 MB in dynamic EAKV Q8_0/Q4_0, and 624 MB in static --cache-type-k q8_0 --cache-type-v q4_0 across all 262,144 tokens) while preserving FP16-equivalent needle retrieval on keys.
8. Zero-Prefill Agentic Loops (HARP-Cache Prefix Tracker + Native /dev/shm Slot Restore + SSE Gateway) (ARCHITECTURE.md §3.8)
   Separates gateway-level prefix tracking from native C++ tensor persistence: kv_cache/harp_cache.py provides the gateway's 32-token FNV-1a Radix Tree & bounded /dev/shm LRU page tracker (and standalone algorithmic reference), while llama-server --slot-save-path /dev/shm --cache-reuse 256 (scripts/prewarm_agent_prefix.py + _ensure_slot0_restored() in gateway/sse_gateway.py) performs native C++ serialization/restoration of the 85 MB Mamba-2 + GQA state in 17.2–18.9 ms.
9. Norm-Preserving Dynamic MoE Top-K Pruning (LLAMA_MOE_TOP_K=4) (ARCHITECTURE.md §3.9)
   • On Nemotron-3.5-30B-A3B (🟢 SAFE): Prunes active routed experts from 6 -> 4 (8 -> 6 total expert-units including the double-width shared expert, -33.3% routed MoE DRAM traffic). Because expert_weights_norm=True (scale=2.5), llama.cpp automatically renormalizes the surviving 4 gates to sum to 1.0 * 2.5.
   • On Qwen3-30B-A3B & Qwen3.5-MoE (🟢 SAFE): Prunes routed experts from 8 -> 4 (-50% routed MoE DRAM traffic, +38.5% decode). Even though the GGUF converter omits qwen3moe.expert_weights_norm, llama.cpp hardcodes norm_w = true (LLM_FFN_SILU, true) inside src/models/qwen3moe.cpp:144 (build_moe_ffn) (weights = ggml_div(ctx0, weights, weights_sum)), automatically renormalizing the surviving 4 experts so $\sum_{i=1}^{4} \tilde{g}_i = 1.0$ without L2 norm collapse! Auto-locked on Dense models.

4. Quickstart & Reproducible Setup

Option A: 1-Command Automated Installer (Starting from Zero)

If you are starting from a fresh Linux machine, install.sh automates the entire stack end-to-end:

1. Runs the hardware readiness audit (scripts/check_cpu_readiness.py), checks/installs system build tools (cmake, g++, openmp), and configures 2MB Transparent HugePages.
2. Clones llama.cpp at the tested commit (d2f83055), applies patches/zen5-nemotron-256k-optimizations.patch, injects the native RouteCrack C++ engine (speculative/cpp/), and compiles llama-server with auto-detected SIMD flags (AVX-512 VNNI+BF16 or AVX2).
3. Downloads NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Q4_0.gguf (17.60 GiB, auto-resumable) from Hugging Face (or links your existing GGUF via --model /path/to/model.gguf).
4. Tokenizes a local coding + <tool_call> corpus against the 131K Nemotron vocabulary and builds the 8-level RouteCrack FNV-64 hash cascade, hot token rules, and macro codebook.
5. Generates auto-tuned ./start_server.sh and ./start_gateway.sh scripts matched to your physical core count and RAM.

# One-liner from scratch:
curl -sL https://raw.githubusercontent.com/synchopate/nemotron-30b-cpu-256k/main/install.sh | bash

# Or clone and use the unified ./modchip CLI:
git clone https://github.com/synchopate/nemotron-30b-cpu-256k.git
cd nemotron-30b-cpu-256k
./modchip setup                       # Or custom corpus: ./modchip setup --corpus /path/to/codebase
./modchip start                       # Starts server (8789) + Web UI & SSE Gateway (8787) in background

# Verify speed & chat immediately (Terminal, Browser Web HUD, Benchmark, or curl):
./modchip chat                        # Interactive terminal chat with live tok/s footer
./modchip bench                       # Automated 2-stage benchmark + shareable CPU scorecard
./modchip web                         # Opens http://127.0.0.1:8787 (Dark-Mode Web Chat + Live Telemetry HUD)
./modchip claude                      # Launches Claude Code CLI wired to local 256K CPU engine

# Or test directly with a 1-line curl smoke test:
./modchip chat -p "Write a cache-aligned C++20 spinlock in 15 lines."

Option B: Manual Step-by-Step Build

Step 1: Enable 2MB Transparent HugePages (Linux Host)

echo always | sudo tee /sys/kernel/mm/transparent_hugepage/enabled
echo defer+madvise | sudo tee /sys/kernel/mm/transparent_hugepage/defrag
sudo cp systemd/thp-llm.conf /etc/tmpfiles.d/thp-llm.conf

Step 2: Build llama.cpp with Zen 5 / AVX-512 Patch & RouteCrack Engine

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
git apply /path/to/nemotron-30b-cpu-256k/patches/zen5-nemotron-256k-optimizations.patch

# Integrate the native RouteCrack + Stockfish-PVS C++ predictor into common/
cp /path/to/nemotron-30b-cpu-256k/speculative/cpp/rc_predictor.h common/
cp /path/to/nemotron-30b-cpu-256k/speculative/cpp/rc_trie_v2.h common/
cp /path/to/nemotron-30b-cpu-256k/speculative/cpp/rc_stockfish.h common/
cp /path/to/nemotron-30b-cpu-256k/speculative/cpp/ngram-map.cpp common/

cmake -B build \
  -DCMAKE_BUILD_TYPE=Release \
  -DGGML_NATIVE=ON \
  -DGGML_AVX512=ON \
  -DGGML_AVX512_VNNI=ON \
  -DGGML_AVX512_VBMI=ON \
  -DGGML_AVX512_BF16=ON \
  -DGGML_OPENMP=ON
cmake --build build --config Release -j $(nproc) --target llama-server

Step 3: Build Native Nemotron RouteCrack Tables (Optional Custom Corpus)

g++ -O3 -std=c++17 speculative/cpp/rc_tokenize_nemotron.cpp -o rc_tokenize_nemotron \
  -I/path/to/llama.cpp/include -I/path/to/llama.cpp/ggml/include \
  -L/path/to/llama.cpp/build/bin -lllama -Wl,-rpath,/path/to/llama.cpp/build/bin
g++ -O3 -std=c++17 speculative/cpp/rc_build_tables.cpp -o rc_build_tables

./rc_tokenize_nemotron ./workspace/models/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Q4_0.gguf \
  train.txt ./workspace/routecrack/corpus_nemotron_train.bin
./rc_build_tables ./workspace/routecrack/corpus_nemotron_train.bin ./workspace/routecrack

Step 4: Launch Production Service & Instant Slot 0 Pre-Warmer

See systemd/nemotron-256k.service and scripts/prewarm_agent_prefix.py for the complete systemd units.

5. Repository Structure

nemotron-30b-cpu-256k/
├── LICENSE                                           # Apache 2.0 License
├── modchip                                           # Unified CLI (setup | start | chat | bench | web | claude | stop)
├── install.sh                                        # 1-Command Zero-to-Running Automated Installer
├── README.md                                         # Executive Overview, Quickstart & Hardware Check
├── ARCHITECTURE.md                                   # Full Technical Paper, Proofs & Kernel Benchmarks
├── FAQ.md                                            # Engineering Trade-Offs, NIAH Retrieval & Xeon/EPYC Scaling
├── RELEASE_NOTES.md                                  # Detailed Release Notes (v2.1.0 Stockfish-PVS v2 & Multi-Platform)
├── patches/
│   └── zen5-nemotron-256k-optimizations.patch        # Unified C/C++ Patch for llama.cpp (ggml-cpu)
├── speculative/                                      # RouteCrack + Stockfish-PVS Draftless Speculative Engine
│   ├── ROUTECRACK_TECHNIQUES.md                      # Master Catalog of 96+ Studied RouteCrack Techniques & PVS v2 Ablations
│   ├── cpp/
│   │   ├── ngram-map.cpp                             # Drop-in llama-server integration (RouteCrack + Stockfish-PVS v2 M=7)
│   │   ├── rc_stockfish.h                            # 1.64 MB L2/L3 Chess-Engine PVS + Zobrist TT (tt5/tt4/tt3) + Countermove
│   │   ├── rc_predictor.h                            # 8-Level Hash Cascade, Online Trie, PPM, Skipgram, Hot Rules, Probes
│   │   ├── rc_trie_v2.h                              # 4M-Node Arena Online Learning Trie (O(1) lookup, LRU eviction)
│   │   ├── rc_tokenize_nemotron.cpp                  # Native GGUF vocabulary corpus tokenizer (131,072 tokens)
│   │   ├── rc_build_tables.cpp                       # Laplace-smoothed cascade, hot rules & macro codebook builder
│   │   ├── test_synergy.cpp                          # 3-Way Held-out benchmark (77,456 tok: Baseline vs PVS v1 vs PVS v2 M=7)
│   │   └── rc_engine.cpp                             # Standalone C++ Speculative Engine with BFS Tree Verification
│   └── tools/                                        # Linear & Hot-Logit Probe trainers and distillation scripts
├── kv_cache/                                         # HARP-Cache (/dev/shm) & EAKV + VPI Asymmetric Quantization
├── gateway/                                          # Agentic SSE Reverse Proxy + Live Web Chat & Telemetry HUD (sse_gateway.py)
├── scripts/
│   ├── chat.py                                       # Zero-dependency Interactive Terminal Chat & 2-Stage Benchmark Scorecard
│   ├── llama_modchip.py                              # "Priiloader for CPU LLM" (GGUF Inspector, Safety Lock & Live SHM TUI)
│   ├── check_cpu_readiness.py                        # Zero-dependency CPU compatibility & tok/s estimator script
│   ├── prewarm_agent_prefix.py                       # 18.9ms Slot 0 disk snapshot restore & prefix pre-warmer
│   ├── build_zen5_llama.sh                           # CMake build script for Zen 5 AVX-512 VNNI
│   └── fill_kv_cache_256k.py                         # 256K context scaling & TTFT stress test
├── systemd/                                          # Production systemd unit & 2MB Transparent HugePages config
├── apple-silicon/                                    # Apple M4 Pro (24 GB) Metal setup, scripts & 128K benchmarks
├── fedora/                                           # Fedora 44 / AMD Ryzen 7 8745HS (GEEKOM A8) setup & web chat script
├── benchmarks/                                       # Detailed KV scaling & Hybrid SSM vs Dense GQA reports
└── reviews/                                          # Microarchitectural audit & 4-Case Kernel Code Auditing evaluation

🔗 Continue Reading:

• ARCHITECTURE.md — Full Technical Paper & Benchmarks
• speculative/ROUTECRACK_TECHNIQUES.md — Master Catalog of 96+ RouteCrack Techniques & Stockfish-PVS v2 Ablations
• FAQ.md — Technical FAQ, Long-Context Retrieval & Intel Xeon / AMD EPYC Scaling
• RELEASE_NOTES.md — Release Notes (v2.1.0)