Clarify installation, performance and compatibility for runtime users

releases/20261022-01/README-20261022-01.md

@@ -1,80 +1,57 @@
----
-license: mit
-tags:
-- llama-cpp
-- rocm
-- amd
-- gfx1201
-- windows
-- inference-engine
-- experimental
----
-
-# AiTable ROCmFPX Windows gfx1201 — 20261022-01
-
-## Pair this runtime with the model
-
-This is a **runtime binary release**, not a model. The tested matching model is
-[Ornith Q6 IO + A8](https://huggingface.co/Roland2531/Ornith-1.5-9B-uncensored-ROCmI4-Q6IO-A8).
-Download its GGUF separately. The model and runtime have independent licenses.
-Other models/formats are not covered by this release's real-model acceptance.
-
-This release is grouped under [releases/20261022-01](https://huggingface.co/Roland2531/llama.cpp-ROCmFPX-gfx1201-Windows-x64/tree/main/releases/20261022-01).
-The previous [20261021-01 runtime](https://huggingface.co/Roland2531/llama.cpp-ROCmFPX-gfx1201-Windows-x64/blob/main/AiTable-ROCmFPX-gfx1201-Windows-x64-20261021-01.zip)
-and its versioned documents remain available without replacement.
-
-## Source and build
-
-- llama.cpp integration target: `9febe63d5a379f8568bae7116a5b2ba844a4a4a6` from [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp/tree/9febe63d5a379f8568bae7116a5b2ba844a4a4a6).
-- ROCmFPX base: `fb08d7cdb67082ddda15a25e64dc7ec413afc41c` from [charlie12345/ROCmFPX](https://github.com/charlie12345/ROCmFPX).
-- Private legacy FPX kernels bridged into the modern backend; Q6 IO + W4A8, MV2, RoPE/FWHT, short prefix checkpoints, CK bridge, native grammar/diagnostics and projector lifecycle retained.
-- Mask writing now follows the actual F16 mask type. Quantized state storage views count logical elements rather than block counts. RAM and independent device snapshot round-trips matched reference logits exactly in the narrow native probe.
-- Windows HIP clang **24.0.0git**, compiler commit `d6f6cb691863e9258034ab48c1dfce21689d5606`; HIP header **7.16.26385**. The local SDK folder was named ROCm10.1-LLVM24; that folder label is not a claim that every component is an official ROCm 10.1 release.
-- Release, target gfx1201, HIP graphs, forced MMQ and all-quant Flash Attention enabled. Native W4A4 disabled. rocBLAS used; hipBLASLt not selected for these tests.
-- Server, CLI, mtmd CLI, completion and bench all use matching new DLLs. New `*-impl.dll` files are required; do not copy only the tiny EXE launchers or retain older EXEs.
-- Full source archives remain local and are not part of this publication. Old `apply-fpx-fixes-20261021-01.py` patches target the old base and must **not** be applied to this integrated release.
-
-## Installation
-
-Extract the complete ZIP. Its top-level path is `Runtimes/llama.cpp/FPX_CHECKPOINT/`.
-For standalone use, keep that directory intact and run from it. For AiTable, stop the service, back up the old directory, then replace the whole backend directory; keep model JSON and GGUF separately.
-All bundled AMD DLLs and rocBLAS/hipBLASLt data must stay with the runtime; retain the adjacent `../.kpack/` AMD sidecar directory as supplied. An installed compatible AMD driver and Microsoft VC++ runtime are still required.
-The backend is validated on Windows x64 / RX 9070 (`gfx1201`). Linux and other GPU architectures were not built/tested.
-
-From cmd.exe, `call configure-fpx-env.cmd` applies the tested environment relative to the runtime directory.
-For a standalone text-only server (replace the model path), use:
-
-```bat
-llama-server.exe -m "MODEL.gguf" --host 127.0.0.1 --port 8080 -c 32768 -ngl 99 -fa on -ctk q8_0 -ctv q8_0 --parallel 1 --cache-ram 1024 --jinja --reasoning off --no-mmproj --temp 1 --top-p 0.95 --top-k 20 --min-p 0 --presence-penalty 1.5 --repeat-penalty 1 --spec-type ngram-mod --spec-ngram-mod-n-match 6 --spec-ngram-mod-n-min 1 --spec-ngram-mod-n-max 4
-```
-
-AiTable already assembles its own environment/arguments from model settings; the script does not alter model files. No credentials are included. Use the service's supported environment-based key configuration before exposing access.
-
-## Observed performance, not a universal speed claim
-
-Same Ornith Q6 IO + A8, no thinking/MTP, ngram-mod 6/4, one slot, Q8 K/V, RAM cache 1024 MiB, 32768 configured context; active tool prompt **3580 tokens**, output limit 2048. Five samples per condition; one service per version, no reverse-order repetitions.
-
-| HTTP metric | Previous installed runtime | This release |
-| --- | ---: | ---: |
-| Cold first nonempty tool argument | 1084.2 ms | 1056.3 ms |
-| Cold prefill | 3596.6 token/s | 3668.7 token/s |
-| Cold generation | 94.0 token/s | 93.8 token/s |
-| Exact-prefix warm first tool argument | 107.2 ms | 132.9 ms |
-
-Warm prompts reused 3576/3580 tokens. This is not time to audible speech.
-The earlier native CLI 3072-token fresh prefill was 845.8 vs 767.2 ms and generation 84.6 vs 86.3 token/s; it is a different workload and must not be substituted for HTTP results.
-
-With the same diagnostic A→A→background B→A→prefix edit→capability-query sequence, single-process sampled peaks were dedicated GPU 6230.3→5513.0 MiB, process private RAM 5297.1→5862.9 MiB, shared GPU 193.1→179.5 MiB. Detailed logs were enabled for this resource sequence. These are observed peaks, not full 32K occupancy or LLM/TTS/Viewer coexistence measurements.
-
-## Acceptance and limitations
-
-- 18 complete current-schema candidates passed JSON Schema shape checks. This does **not** establish semantic tool correctness: both runtimes produced speech for silence/nod requests and rewrote exact requested lines. Some host replays were rejected or blocked by unavailable live embedding dependencies. Disabling ngram alone did not fix the two-line instruction case.
-- Both variants recovered the original synthetic source identifier across A→background B→A. This release handled one early-prefix identifier update; the old runtime mixed the new identifier's actor attribution with obsolete data. These are limited cases, not general cache correctness proof.
-- Both variants performed the actual read-only capability query and then used a host-composed synthetic result; no real Viewer was driven.
-- Client disconnect returned the native HTTP slot to idle in about 102 ms for this release (old about 115 ms); subsequent reply did not continue the cancelled story. This is not audio/Viewer cancellation acceptance.
-- Native quantization conversion/decode matched old bytes in 135 paired cases; selected source/parser/Jinja/grammar CPU tests and native model probes passed. All tested core/AMD loaded paths were checked and AMD dependency hashes matched across versions.
-- No observed driver-timeout events during the isolated HTTP run. All test services closed.
-- Vision CLI/projector lifecycle support is present, but real vision inference/Viewer/audio, a filled 32768-token context, coexistence, exhaustive long-loop tests, Linux and other GPU architectures were not accepted in this release.
-- Alternate FPX weight/KV formats only retain compilation/conversion coverage; the previous Turbo KV fused vector fast path was not ported. No universal regression-free claim.
-
-The package includes llama.cpp MIT license and the bundled third-party license directory. Review those notices for individual AMD components.
+# ROCmFPX Runtime 20261022-01
+
+## Compatibility
+
+Windows x64 community build for AMD Radeon RX 9070 (`gfx1201`), tested with [Ornith Q6 IO + A8](https://huggingface.co/Roland2531/Ornith-1.5-9B-uncensored-ROCmI4-Q6IO-A8). Download the model GGUF separately. Other model formats, GPUs and operating systems are unverified.
+
+This release integrates llama.cpp `9febe63d5a379f8568bae7116a5b2ba844a4a4a6` and ROCmFPX base `fb08d7cdb67082ddda15a25e64dc7ec413afc41c`. Build tools: AMD clang 24.0.0git (`d6f6cb691863e9258034ab48c1dfce21689d5606`), HIP headers 7.16.26385, MSVC. HIP graphs, forced MMQ and all-quant Flash Attention are enabled; native W4A4 is disabled.
+
+Included features: Q6 IO/W4A8 kernels, MV2, RoPE/FWHT, CK attention bridge, rocBLAS, short-prefix checkpoints, native tool grammar/diagnostics and projector lifecycle support. Fixes cover typed F16 mask writes and quantized snapshot element counts. Server, CLI, mtmd CLI, completion and bench use matching shared libraries.
+
+## Installation
+
+1. Install a compatible AMD driver and the Microsoft Visual C++ x64 runtime.
+2. Extract the complete ZIP. Keep `Runtimes/llama.cpp/FPX_CHECKPOINT/` and adjacent `Runtimes/llama.cpp/.kpack/` intact, including AMD DLLs, library data and license files.
+3. For AiTable, stop its service, back up the previous backend directory, then replace the complete backend. Model GGUF/JSON files remain separate.
+4. For standalone use, open cmd.exe in `FPX_CHECKPOINT` and run `call configure-fpx-env.cmd` before launching a tool.
+
+The tiny executable launchers require the matching `*-impl.dll` files. Do not copy only the EXE or reuse old DLLs. The previous version's source patch scripts do not apply to this integration.
+
+## Text-only server example
+
+Replace `MODEL.gguf` with your model path:
+
+```bat
+call configure-fpx-env.cmd
+llama-server.exe -m "MODEL.gguf" --host 127.0.0.1 --port 8080 -c 32768 -ngl 99 -fa on -ctk q8_0 -ctv q8_0 --parallel 1 --cache-ram 1024 --jinja --reasoning off --no-mmproj --temp 1 --top-p 0.95 --top-k 20 --min-p 0 --presence-penalty 1.5 --repeat-penalty 1 --spec-type ngram-mod --spec-ngram-mod-n-match 6 --spec-ngram-mod-n-min 1 --spec-ngram-mod-n-max 4
+```
+
+This binds loopback only, disables thinking/MTP and uses ngram speculation. The preset enables CK attention; earlier tests observed input-dependent output differences with CK enabled, so correctness-sensitive workloads require their own checks. `--cache-ram 1024` caps host-RAM prompt cache in MiB; it does not reserve VRAM. AiTable supplies its own environment and arguments from model settings.
+
+## Performance
+
+Same Ornith Q6 IO + A8, no thinking/MTP, ngram-mod 6/4, Q8 K/V, one slot, RAM cache 1024 MiB, configured context 32768, output limit 2048. Active tool prompt: **3580 tokens**. Five measurements per condition; single service per version, without reverse-order repetitions.
+
+| HTTP metric | 20261021-01 baseline | 20261022-01 |
+| --- | ---: | ---: |
+| Cold first nonempty tool argument | 1084.2 ms | 1056.3 ms |
+| Cold prefill | 3596.6 token/s | 3668.7 token/s |
+| Cold generation | 94.0 token/s | 93.8 token/s |
+| Exact-prefix warm first tool argument | 107.2 ms | 132.9 ms |
+
+Warm requests reused 3576/3580 tokens. Tool-argument latency is not time to audible speech. Small differences do not establish a general speedup.
+
+For a matched cache-switch/capability-query diagnostic workload, sampled dedicated GPU peaks were 6230.3 vs 5513.0 MiB; process private RAM 5297.1 vs 5862.9 MiB; shared GPU memory 193.1 vs 179.5 MiB. These single-process peaks do not represent a filled 32K context or simultaneous LLM/TTS/STT/Viewer use.
+
+Separate native CLI tests with fresh 3072-token contexts measured prefill 845.8 vs 767.2 ms and generation 84.6 vs 86.3 token/s. These use a different workload and do not replace the HTTP figures.
+
+## Validation and known limits
+
+- Native quantization encode/decode matched the previous build in 135 paired cases. Narrow RAM/device snapshot round-trips matched reference logits exactly. Selected CPU parser/grammar/Jinja tests and real-model HTTP checks passed.
+- Eighteen tool candidates passed JSON Schema structure checks. Both versions still failed some silence, exact wording and tool instructions; disabling ngram alone did not resolve the tested two-line case. Structured output is not a guarantee of semantic accuracy.
+- Tested A→background B→A requests recovered the source identifier. The newer build also handled one early-prefix identifier update; these cases are not proof for all conversations.
+- Native HTTP disconnect returned the slot to idle in about 102 ms; the following request did not continue the cancelled story. This does not validate application audio/Viewer cancellation.
+- Projector CLI/lifecycle support is present, but real vision, Viewer/audio integration, filled-context coexistence, exhaustive long-loop workloads, Linux and other GPU architectures remain unverified.
+- Alternate FPX weight/KV formats have limited compilation/conversion coverage. The previous Turbo KV fused vector fast path is not included.
+
+Core code uses MIT; preserve `LICENSE-llama.cpp.txt` and bundled dependency licenses. Model licensing is separate. [Previous runtime](https://huggingface.co/Roland2531/llama.cpp-ROCmFPX-gfx1201-Windows-x64/tree/main/releases/20261021-01).

releases/20261022-01/README.md

@@ -1,9 +1,14 @@
-# FPX Runtime 20261022-01 — integrated version
+# ROCmFPX Runtime 20261022-01
+
+Windows x64 / RX 9070 (`gfx1201`) runtime with newer llama.cpp integration.
+
+## Downloads
 
 - [Complete runtime ZIP](AiTable-ROCmFPX-gfx1201-Windows-x64-20261022-01.zip)
-- [Build, usage, paired model, benchmarks and limitations](README-20261022-01.md)
+- [Installation, parameters, performance and limitations](README-20261022-01.md)
 - [Build manifest](build-manifest-20261022-01.json)
 - [SHA-256 checksums](SHA256SUMS-20261022-01.txt)
-- [Previous version](../20261021-01/)
+- [Matching Ornith Q6 IO + A8 model](https://huggingface.co/Roland2531/Ornith-1.5-9B-uncensored-ROCmI4-Q6IO-A8)
+- [Previous runtime](../20261021-01/)
 
-On 2026-10-22 the older root files were relocated into their own version folder without changing their bytes. Version-specific notes remain as published; use these indexes for current paths. Full source archives and private test inputs are not uploaded.
+Extract the complete package. Keep `Runtimes/llama.cpp/FPX_CHECKPOINT/` and its adjacent `.kpack/` data intact. The previous release's patch scripts are not compatible with this integrated source base.

releases/20261022-01/SHA256SUMS-20261022-01.txt

@@ -1,3 +1,3 @@
-06560aac536cd2050547e39957a6f10ea9d46a470fe6d186a94fe53e0e0481af  AiTable-ROCmFPX-gfx1201-Windows-x64-20261022-01.zip
-5dc473ae7a5504021041e3f207d549a8edac6625a3c77e60c12a8a95a5768872  build-manifest-20261022-01.json
-8781b774d8c91ba13d62100503bd357df0b6c745f5c1817d8c09423acdb37d01  README-20261022-01.md
+06560aac536cd2050547e39957a6f10ea9d46a470fe6d186a94fe53e0e0481af  AiTable-ROCmFPX-gfx1201-Windows-x64-20261022-01.zip
+0fee899cc9141e97d37363978862c76846a1664868c8d21ddcc83243efe90fb0  README-20261022-01.md
+5dc473ae7a5504021041e3f207d549a8edac6625a3c77e60c12a8a95a5768872  build-manifest-20261022-01.json