Skip to content

Latest commit

 

History

History
61 lines (43 loc) · 2.41 KB

RELEASE_NOTES.md

File metadata and controls

61 lines (43 loc) · 2.41 KB

v0.1.0-rdna4-seed — RX 9070 16GB gfx1201 Vulkan Seed

First RDNA4 lane for 1×-RDNA16. Clone of [5p00kyy/benchmark protocol](https://github.com/5p00kyy/benchmark protocol) — same protocol, new gfx1201 seed on Windows Vulkan.

Seed Verified 2026-09-01

Field Value
GPU AMD Radeon RX 9070 16GB RDNA4 gfx1201, 640GB/s
Host Ryzen 7 7800X3D, 64GB DDR4, Windows 11 Vulkan
llama.cpp b10712 daef7b687, Vulkan ggml-vulkan.dll, flash_attn on
Model Qwen3.5-9B-UD-Q4_K_XL.gguf 5.55GB Q4_K, 8.95B 32L
Runtime 131072 ctx q8_0 KV, b512/ub256, reasoning on, parallel=1, 22 MCP tools
VRAM 8.8GB @65k q4 -> 10.0GB @131k q8
Hybrid Muse Spark (orchestrator) + Qwen 3.5-9B local (explore/general/sdd-*) @70-80 tok/s

Evidence — Bench Sweep (don't guess, sweep)

\
llama-bench -m gguf/Qwen3.5-9B-UD-Q4_K_XL.gguf -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 -p 2048 -n 128 -b 512,1024,2048,4096 -ub 64,128,256,512 \\

  • Prompt 2048: b512/ub512 2205.5 tok/s (top), b1024/ub512 2204.3, b512/ub256 2153.4
  • Gen 128: ~86-87 tok/s all
  • Long receipt: 8192/512 q8 -> 5350B JSON, no OOM, 5.3s — public recommendation ready

Raw receipts: \�ench/receipts/q8-sweep-2026-09-01.json\ (21298B, 16 combos) + \q8-long.json\

Per protocol: short-only fast is short-evidence. Long receipt satisfies public lane.

Heavy Validation — 15min @100%

  • heavy-15min (in-memory, no SQLAlchemy): 62/62 tests in 6.24s, 49KB app (store 12k, main 18k), RLock, TTL, Thread-safe, Docker
  • kvstore (earlier): 17/17 in 2.18s
  • 25k prompt: 14.45s @1764 tok/s, 168 gen @77.5 tok/s, Compute_0 90% 14s

GPU sustained 70-80 tok/s, 15min @100% verified via Task Manager Compute_0 + llama log.

Presets


  • un-qwen3.5-9b-server-opencode.bat\ — 65k q4, 8.8GB, daily

  • un-qwen3.5-9b-server-opencode-131k-q8.bat\ — 131k q8, b512/ub256, 10GB, heavy

Hybrid Workflow

\
User -> Muse Spark (propose/spec/design/verify) -> Qwen local (explore/tasks/apply) @70 tok/s -> Muse review \
Opencode: \llamacpp/http://localhost:8080/v1\ + Bearer anything, agents explore/general/sdd-* -> qwen3.5-9b

Assets

  • \�ench/receipts/q8-sweep-2026-09-01.json\ + \q8-long.json\
  • \heavy-15min\ 62 tests
  • Docs: \docs/rx-9070-baseline.md, \docs/benchmark-protocol.md\

Upstream: https://github.com/5p00kyy/benchmark protocol


Seed by albertgracia + Muse Spark + Qwen3.5-9B local (RX 9070). Measure, don't assume.