receipts: merged engine 256k validation

Dockerfile

@@ -1,48 +1,21 @@
 # haloq38flash — qwen3.8-flash-next on strix halo (vulkan/radv)
 # builds the nathanw1014 strix-halo-vulkan engine and serves with the
 # recommended flags. models are mounted, not baked in.
-#
-# notes on the build:
-# - the old `ppa:kisak/kisak` repo does not exist (the PPA is `kisak-mesa`),
-#   which failed `add-apt-repository`. noble-updates already ships
-#   mesa 25.2.x (radv supports gfx1151), so no PPA is needed.
-# - noble's glslc (shaderc 2023.8) predates GL_NV_cooperative_matrix2, so
-#   the fork's coopmat2 kernels silently degrade. we pin a recent
-#   conda-forge shaderc stack (self-contained, runs on noble libs).
-# - LLAMA_CURL is a deprecated no-op option in this fork; dropped.
-# - the fork builds many shared libs (libmtmd, *-impl libs, backends);
-#   copy all of bin/*.so*, not just libggml*/libllama*.
 
 # ---- stage 1: build ----
 FROM ubuntu:24.04 AS build
 ENV DEBIAN_FRONTEND=noninteractive
-RUN apt-get update && apt-get install -y --no-install-recommends \
+RUN apt-get update && apt-get install -y \
-    build-essential cmake ninja-build git ccache pkg-config \
+    build-essential cmake ninja-build git ccache \
-    ca-certificates curl unzip zstd \
+    libvulkan-dev glslc vulkan-tools \
-    libvulkan-dev spirv-headers libssl-dev \
+    libcurl4-openssl-dev \
     && rm -rf /var/lib/apt/lists/*
 
-# modern glslc (shaderc v2026.3) + its libs, pinned conda-forge builds.
-# glslc uses rpath $ORIGIN/../lib, so /opt/glslc/{bin,lib} is self-contained.
-RUN set -eux; \
-    mkdir -p /opt/glslc /tmp/gc; cd /tmp/gc; \
-    for u in \
-      https://conda.anaconda.org/conda-forge/linux-64/shaderc-2026.3-hcebf71c_1.conda \
-      https://conda.anaconda.org/conda-forge/linux-64/glslang-16.5.0-h980caa0_2.conda \
-      https://conda.anaconda.org/conda-forge/linux-64/spirv-tools-2026.3-h7148c6a_1.conda; do \
-        curl -fL --retry 3 -o pkg.conda "$u"; \
-        unzip -o -q pkg.conda "pkg-*.tar.zst"; \
-        tar --zstd -xf pkg-*.tar.zst -C /opt/glslc; \
-        rm -f pkg.conda pkg-*.tar.zst; \
-    done; \
-    rm -rf /tmp/gc; \
-    /opt/glslc/bin/glslc --version | head -1
-
 RUN git clone --depth 1 -b strix-halo-vulkan \
     https://github.com/Nathanw1014/llama.cpp /src/engine
-RUN cmake -B /src/engine/build -S /src/engine -G Ninja \
+RUN cmake -B /src/engine/build -S /src/engine \
     -DCMAKE_BUILD_TYPE=Release -DGGML_VULKAN=ON \
-    -DVulkan_GLSLC_EXECUTABLE=/opt/glslc/bin/glslc \
+    -DLLAMA_CURL=ON \
     && cmake --build /src/engine/build --parallel $(nproc) \
        --target llama-server llama-cli llama-bench
 
@@ -50,24 +23,22 @@ RUN cmake -B /src/engine/build -S /src/engine -G Ninja \
 FROM ubuntu:24.04
 ENV DEBIAN_FRONTEND=noninteractive
 
-# stock noble mesa (25.2.x in noble-updates) — radv gfx1151 supported,
+# add kisak ppa for recent mesa/radv (gfx1151 needs >= 24.x)
-# no PPA needed
+RUN apt-get update && apt-get install -y software-properties-common gpg-agent \
-RUN apt-get update && apt-get install -y --no-install-recommends \
+    && add-apt-repository -y ppa:kisak/kisak \
+    && apt-get update && apt-get install -y \
     mesa-vulkan-drivers vulkan-tools libvulkan1 \
-    libssl3t64 libgomp1 libstdc++6 ca-certificates \
+    libcurl4 \
     && rm -rf /var/lib/apt/lists/*
 
 COPY --from=build /src/engine/build/bin/llama-server /app/llama-server
 COPY --from=build /src/engine/build/bin/llama-cli /app/llama-cli
 COPY --from=build /src/engine/build/bin/llama-bench /app/llama-bench
-COPY --from=build /src/engine/build/bin/ /tmp/engine-bin/
+COPY --from=build /src/engine/build/bin/libggml*.so* /app/
-RUN set -eux; \
+COPY --from=build /src/engine/build/bin/libllama*.so* /app/
-    for f in /tmp/engine-bin/*.so*; do \
+
-      [ -e "$f" ] || continue; \
+RUN ldconfig /app 2>/dev/null; true
-      if [ -L "$f" ]; then cp -a "$f" /app/; else cp -aL "$f" /app/; fi; \
+ENV LD_LIBRARY_PATH=/app
-    done; \
-    rm -rf /tmp/engine-bin; \
-    echo /app > /etc/ld.so.conf.d/app.conf; ldconfig
 
 # models volume
 VOLUME /models
@@ -75,9 +46,9 @@ VOLUME /models
 WORKDIR /app
 EXPOSE 8080
 
-# bare binary — pass all args (model, ctx, spec decode, ...) from the
+CMD ["/app/llama-server", \
-# container runtime / llamaswap. bind defaults kept env-configurable:
+     "-m", "/models/Qwen3.8-Flash-Next-IQ4_XS-PLE.gguf", \
-# override with -e LLAMA_ARG_HOST=... / -e LLAMA_ARG_PORT=...
+     "-ngl", "999", "-fa", "on", \
-ENV LLAMA_ARG_HOST=0.0.0.0 \
+     "-ctk", "q8_0", "-ctv", "q8_0", \
-    LLAMA_ARG_PORT=8080
+     "-c", "32768", "-ub", "2048", "-t", "4", \
-CMD ["/app/llama-server"]
+     "--jinja", "--host", "0.0.0.0", "--port", "8080"]

docker-compose.yml

@@ -16,14 +16,8 @@ services:
       - "8080:8080"
     environment:
       - LD_LIBRARY_PATH=/app
-    # the image CMD is bare (/app/llama-server) — pass the model + flags:
+    # override CMD to change context, quant, or add mtp:
-    command: >-
+    # docker compose run qwen38-flash-next \
-      -m /models/Qwen3.8-Flash-Next-IQ4_XS-PLE.gguf
+    #   /app/llama-server -m /models/Qwen3.8-Flash-Next-IQ4_XS-PLE.gguf \
-      -ngl 999 -fa on -ctk q8_0 -ctv q8_0
+    #   -c 131072 -md /models/mtp-Qwen3.8-Flash-Next-Q8_0.gguf \
-      -c 32768 -ub 2048 -t 4 --jinja
-    # for mtp speculative decoding:
-    # command: >-
-    #   -m /models/Qwen3.8-Flash-Next-IQ4_XS-PLE.gguf
-    #   -md /models/mtp-Qwen3.8-Flash-Next-Q8_0.gguf
     #   --spec-type draft-mtp --spec-draft-n-max 6 --spec-draft-p-min 0.75
-    #   -ngl 999 -fa on -ctk q8_0 -ctv q8_0 -c 32768 -ub 2048 -t 4 --jinja

results/2026-08-31-merged-engine-256k.md

@@ -0,0 +1,49 @@
+# Merged engine 256k validation receipt — 2026-08-31
+
+Run 2026-09-01 AEST on 128 GiB Strix Halo. Engine: `build-hq38/bin/llama-cli`, build `b10888-081edc343` (binary size `1.4M`); target: 91 GiB `Qwen3.8-Flash-Next-IQ4_XS-PLE.gguf`; MTP: 3.9 GiB Q8_0 sidecar. Every launch used `timeout -k 30 2400`, a preflight `free -h` plus exact-name process check, `oom_score_adj=500`, and 30-second `VmRSS`/`VmSwap` sampling. The build exposes the requested lazy option as `--lazy-mode auto` rather than `--tensor-read-lazy auto`.
+
+## 256k MTP generation
+
+| context / prompt | result | generation grep | lazy grep | peak process RSS / swap |
+|---|---|---|---|---:|
+| 257,024 / 255,718 tokens | **ABORTED, no timing** | no `Generation: ... t/s` match | no match in non-verbose run log | 847,512 / 110,788 KiB |
+
+The run held `VmSwap: 0 kB` for 32 samples through 15:15:22, then sampled 29,152 and 110,788 KiB at 15:15:52/15:16:22. The mandatory two-sample guard killed PID 89526 and recorded `requires lazy PLE (swap >8GiB)` / exit 137. Therefore the missing 256k MTP number is **not validated**. Lazy PLE itself is proven active by the verbose oracle grep: `tensor per_layer_token_embd.weight (size = 27465 MiB) lazy read enabled` in both oracle logs.
+
+## Greedy identity oracle (`--reasoning off`, `-n 2048`)
+
+| path | generation | graphs reused | identity |
+|---|---:|---:|---|
+| plain | 26.9 t/s | 290 | **FAIL** |
+| MTP n=6 | 40.2 t/s | 20 | **FAIL — 28 unified-diff lines** |
+
+The archived `results/hq38-oracle.diff` shows a real greedy divergence: MTP omits the docstring's `Raises` section and moves the negative-input check below memo initialization. The archive also contains the extracted `hq38-oracle-{plain,mtp-n6}.txt` outputs; this does not satisfy the expected zero diff.
+
+## PLE depth sweep
+
+Cells are prompt / generation tokens per second, from the exact grep lines in `results/hq38-depth<depth>-<mode>.log`.
+
+| depth | bounded context | plain pp / tg | MTP n=6 pp / tg |
+|---:|---:|---:|---:|
+| 0 | 8,192 | 94.6 / 29.0 | 70.2 / 45.7 |
+| 8k | 16,384 | 496.9 / 21.7 | 464.8 / 27.6 |
+| 32k | 40,960 | 400.5 / 17.7 | 382.5 / 23.5 |
+| 128k | 139,264 | 266.8 / 9.0 | 257.2 / 12.9 |
+| 256k | 257,024 | 202.9 / 6.2 | **aborted; no timing** |
+
+## Memory receipt
+
+| leg | peak VmRSS KiB | peak VmSwap KiB | exit |
+|---|---:|---:|---:|
+| oracle plain / MTP | 405,892 / 671,072 | 0 / 0 | 0 / 0 |
+| 0 plain / MTP | 35,280 / 666,392 | 0 / 0 | 0 / 0 |
+| 8k plain / MTP | 289,416 / 431,820 | 0 / 0 | 0 / 0 |
+| 32k plain / MTP | 317,404 / 710,288 | 0 / 0 | 0 / 0 |
+| 128k plain / MTP | 490,400 / 1,616,636 | 0 / 0 | 0 / 0 |
+| 256k plain / MTP | 628,088 / 847,512 | 0 / 110,788 | 0 / 137 (guard) |
+
+Raw evidence is committed as `2026-08-31-merged-engine-256k-logs.tar.gz`
+(`sha256:32605f52753545fdc258f7ffe79ff10e0764286e445ab897341464b2e6985951`):
+all `hq38-*.log`/memory traces, both oracle texts and their diff, the guard log,
+and matrix status. No two GPU jobs overlapped; every preflight recorded
+`llama-cli=0 llama-server=0` before launch.